Data query method and system for three-dimensional gaussian splash scene and electronic equipment

Abstract

The application provides a data query method and system for a three-dimensional Gaussian splash scene and electronic equipment, which can establish a space level and a unique main attribute for a three-dimensional Gaussian primitive, perform global merging in a space node according to a cross-node boundary edge to obtain atomic Gaussian clusters that do not overlap with each other, then generate semantic instances with stable identities based on the atomic Gaussian clusters, further generate immutable scene snapshots that can aggregate multiple atomic Gaussian clusters, and establish geometric evidence constraint conditions corresponding to each semantic level in the scene snapshots; when querying, corresponding geometric and semantic, only semantic or degraded results are returned according to the geometric evidence and resource budget. The method can reduce cross-block repetition, avoid incorrect semantic inheritance, and reduce irrelevant geometric transmission and memory occupation required for fine-grained semantic query.

Description

Technical Field

[0001] This invention relates to the field of data querying of semantic objects in three-dimensional Gaussian scenes, and in particular to a data querying method, system and electronic device for three-dimensional Gaussian splash scenes. Background Technology

[0002] 3D Gaussian splashing technology, with its high rendering performance, has been widely used in large-scale 3D scene applications such as digital twins and autonomous navigation. Existing technologies typically employ spatial partitioning, geometric detail hierarchy, and frustum culling to improve the rendering efficiency of large scenes. Meanwhile, techniques have emerged that attach category probabilities and semantic features to Gaussian primitives to achieve scene semantic retrieval.

[0003] However, existing technologies still have significant drawbacks: First, spatial block-based independent clustering can easily lead to repeated clustering of cross-block objects or semantic breaks, and semantic objects lack globally unique geometric support; second, semantic organization often adopts general parent-child relationships, which cannot distinguish different relationship types such as category inheritance, composition, attachment, and space occupation, and is prone to erroneous confidence inheritance and relationship conflicts; third, the geometric detail level is forcibly bound to the semantic granularity, and querying fine-grained semantics requires synchronous loading of fine geometry, resulting in redundant bandwidth and memory overhead; fourth, progressive queries are not bound to the scene version, and cross-version pagination is prone to result duplication and omission, and state consistency is difficult to guarantee. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a data query method, system, and electronic device for 3D Gaussian splash scenes. This method establishes spatial hierarchy and unique master affiliation for 3D Gaussian primitives, performs global merging within spatial nodes based on cross-node boundary edges to obtain non-overlapping atomic Gaussian clusters, then generates semantic instances with stable identities based on these atomic Gaussian clusters, and further generates an immutable scene snapshot that can aggregate multiple atomic Gaussian clusters. Geometric evidence constraints corresponding to each semantic level are established within this scene snapshot. During querying, the corresponding geometric and semantic, semantic-only, or downgraded results are returned based on geometric evidence and resource budget. This method can reduce cross-block duplication, avoid erroneous semantic inheritance, and reduce irrelevant geometric transfers and GPU memory usage required for fine-grained semantic queries.

[0005] In a first aspect, embodiments of the present invention provide a data query method for a three-dimensional Gaussian splash scene, the method comprising: A spatial hierarchy tree is constructed using scene location data in the target 3D Gaussian splash scene. Based on the spatial hierarchy tree, a unique master-belonging leaf node is determined for each Gaussian record data in the target 3D Gaussian splash scene. After clustering and merging all Gaussian record data using the master-belonging leaf node, an atomic Gaussian cluster is generated. Based on the semantic instance graph generated by the atomic Gaussian cluster, geometric level payloads and semantic level data are constructed respectively. The intersection relationship, semantic content, instance graph and level mapping corresponding to the geometric level payload and semantic level data are integrated into a scene snapshot, and geometric evidence constraints corresponding to each semantic level are constructed. After parsing the query request, a query plan is generated that is bound to the scene snapshot. Under the constraint of geometric evidence, the query results at the geometric and semantic levels are obtained from the scene snapshot according to the query plan.

[0006] Optionally, the generation process of atomic Gaussian clusters includes: For the geometric properties and semantic data of the three-dimensional Gaussian primitives in the target three-dimensional Gaussian splash scene, after verification and normalization processing, multiple Gaussian record data containing calibration category probability, mass weight and bounding box are generated. The target 3D Gaussian splash scene is recursively divided into spatial nodes according to the number of Gaussians and spatial size within the master-owned leaf node. The dividing surface is selected at the mass-weighted median of the longest axis of the node, and the leaf node where the center of each Gaussian record data is located is taken as the unique master-owned leaf node. Within each primary leaf node, the connection weights of candidate pairs are calculated through nearest neighbor search. Then, all Gaussian record data are clustered and merged according to nodes and boundaries using the connection weights to obtain the atomic Gaussian clusters corresponding to each Gaussian record data.

[0007] Optionally, the process of generating the semantic instance graph includes: For the category probabilities of Gaussian record data within an atomic Gaussian cluster, a weighted aggregation is performed according to a preset quality weight to generate corresponding fine-grained semantic instances. Among them, the fine-grained semantic instances adopt a mechanism that separates stable identity identifiers from immutable revision identifiers, so that modifying the semantics only generates new revisions without changing the identity. After aggregating multiple sub-instances in the fine-grained semantic instance according to the execution order corresponding to the semantic level, a virtual aggregated parent instance without independent geometric support is generated. Generate a semantic instance graph corresponding to the atomic Gaussian cluster based on one or more directed relationships among composition, attachment, location, and occupancy between each child instance domain and the virtual aggregate parent instance.

[0008] Optionally, the process of acquiring a scene snapshot includes: By utilizing the intersection relationships, semantic content, instance graphs and hierarchical mappings corresponding to geometric level loads and semantic level data, we construct spatial indexes that support range queries and view frustum queries, semantic inverted indexes that point from categories to instances, instance graph indexes that store type relationships, and mapping indexes that represent detail levels. After pointing the spatial index, semantic inverted index, instance graph index and mapping index to the same snapshot identifier, the instance revision set and load list corresponding to the atomic Gaussian cluster are obtained by using geometric level loads and semantic level data. A scene snapshot of an atomic Gaussian cluster is constructed using the same snapshot identifier, instance revision set, and load list.

[0009] Optional, the process of constructing geometric evidence constraints includes: Obtain the geometric detail level load corresponding to the semantic instance graph based on the semantic level data, and record the original support weight corresponding to the semantic instance graph through the geometric detail level load; Calculate the geometric evidence score corresponding to the original support weights using the projected pixel area of ​​the semantic instance graph in the current view; Geometric evidence constraints are constructed for each semantic level in the semantic instance graph using geometric evidence scores.

[0010] Optional, the query plan generation process includes: After obtaining the query request, the bounding box parameter, radius parameter, and view frustum parameter in the query request are converted to the scene coordinate system corresponding to the scene snapshot; Construct spatial transformation conditions, semantic candidate conditions, and task constraints corresponding to the query request based on the scene coordinate system; A unique query signature is generated using the snapshot identifier, spatial transformation conditions, semantic candidate conditions, and task constraints of the scene snapshot, and a query plan for the scene snapshot is generated based on the query signature.

[0011] Optionally, the process of obtaining query results includes: The spatial index and semantic index corresponding to the query plan are obtained from the scene snapshot using geometric level load and semantic level data, respectively. The candidate instance set is obtained by using the intersection of spatial index and semantic index, and the combination result of geometric level, semantic level and return pattern of each instance in the candidate instance set is obtained; The query results corresponding to the query request are generated by combining the results that satisfy the geometric evidence constraints.

[0012] Optionally, the method further includes: Real-time acquisition of the data space occupied by the query results, and generation of the return batch corresponding to the query request using the data space occupied; Obtain the query signature of the query request, and generate a token verification result for the query signature using the query identifier and snapshot identifier corresponding to the returned batch; Update the query results based on the token verification results.

[0013] Secondly, the present invention provides a data query system for three-dimensional Gaussian splashing scenes, the system comprising: Atomic Gaussian Cluster Generation Module: This module is used to construct a spatial hierarchy tree using scene location data in the target 3D Gaussian splash scene. Based on the spatial hierarchy tree, it determines a unique master-belonging leaf node for each Gaussian record data in the target 3D Gaussian splash scene. Then, it uses the master-belonging leaf node to perform clustering and merging processing on all Gaussian record data to generate atomic Gaussian clusters. The constraint generation module is used to construct geometric level loads and semantic level data based on the semantic instance graph generated by atomic Gaussian clusters. It integrates the intersection relationship, semantic content, instance graph and hierarchical mapping corresponding to the geometric level load and semantic level data into a scene snapshot, and constructs the geometric evidence constraints corresponding to each semantic level. The query result acquisition module is used to parse the query request and generate a query plan that is bound to the scene snapshot. Under the geometric evidence constraint, the query plan is used to obtain the geometric and semantic query results from the scene snapshot.

[0014] Thirdly, embodiments of the present invention also provide an electronic device, which includes a processor and a memory. The memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the steps of the data query method for a three-dimensional Gaussian splash scene provided in the first aspect.

[0015] This invention provides a data query method, system, and electronic device for 3D Gaussian splash scenes. In the process of querying semantic objects in a 3D Gaussian scene, the method first constructs a spatial hierarchy tree using scene location data from the target 3D Gaussian splash scene. Based on the spatial hierarchy tree, a unique master-belonging leaf node is determined for each Gaussian record data in the target 3D Gaussian splash scene. Then, the master-belonging leaf node is used to cluster and merge all Gaussian record data to generate atomic Gaussian clusters. Next, based on the semantic instance graph generated from the atomic Gaussian clusters, geometric level payloads and semantic level data are constructed respectively. The intersection relationships, semantic content, instance graphs, and hierarchy mappings corresponding to the geometric level payloads and semantic level data are integrated into a scene snapshot, and geometric evidence constraints corresponding to each semantic level are constructed. Finally, after parsing the query request, a query plan bound to the scene snapshot is generated. Under the geometric evidence constraints, the query results at the geometric and semantic levels are obtained from the scene snapshot according to the query plan. This solution can reduce cross-block duplication, avoid erroneous semantic inheritance, and reduce the irrelevant geometric transfer and memory usage required for fine-grained semantic queries.

[0016] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a data query method for a three-dimensional Gaussian splash scene provided in an embodiment of the present invention; Figure 2 A flowchart illustrating another data query method for a three-dimensional Gaussian splash scene provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a data query system for a three-dimensional Gaussian splash scene provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0020] icon: 100 - Atom Gaussian Cluster Generation Module; 200 - Constraint Condition Generation Module; 300 - Query Result Acquisition Module; 101 - Processor; 102 - Memory; 103 - Bus; 104 - Communication interface. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To facilitate understanding of this embodiment, a data query method for a three-dimensional Gaussian splash scene disclosed in this embodiment of the invention will first be introduced. The method is as follows: Figure 1 As shown, it includes: Step S101: Construct a spatial hierarchy tree using scene location data in the target 3D Gaussian splash scene. Based on the spatial hierarchy tree, determine a unique master-belonging leaf node for each Gaussian record data in the target 3D Gaussian splash scene. Then, use the master-belonging leaf node to perform clustering and merging processing on all Gaussian record data to generate an atomic Gaussian cluster.

[0023] All Gaussian primitive data of the target 3D Gaussian splash scene are acquired. Based on the spatial distribution characteristics of Gaussians and preset node capacity thresholds, spatial size thresholds, and visual error thresholds, the data is recursively segmented along the longest spatial axis according to the quality-weighted median position, constructing a recursively searchable spatial hierarchy tree. The leaf node where the Gaussian center coordinates are located is taken as its unique master-owned node, and the complete Gaussian data is written into the ownership set of the master-owned node. For Gaussians that cross node boundaries, only a lightweight reference identifier is registered in the adjacent nodes, without copying the complete geometric and semantic data, thus structurally eliminating the problem of duplicate ownership across blocks. First, candidate Gaussian pairs are searched within each leaf node using the nearest neighbor index, and multi-dimensional connection weights such as spatial distance, bounding box overlap, normal consistency, and semantic similarity are calculated. Temporary clusters within the nodes are generated by merging according to thresholds. Then, the connection features are recalculated only for candidate Gaussian pairs at the boundaries of adjacent nodes, and cross-block merging is performed through a global disjoint-set data structure. Finally, globally unique, non-overlapping atomic Gaussian clusters with a determinable processing order are obtained, providing the smallest geometric support unit without repetition or breakage for subsequent semantic organization.

[0024] Step S102: Construct geometric level payloads and semantic level data based on the semantic instance graph generated by the atomic Gaussian clusters. Integrate the intersection relationships, semantic content, instance graphs and level mappings corresponding to the geometric level payloads and semantic level data into a scene snapshot, and construct the geometric evidence constraints corresponding to each semantic level.

[0025] Using atomic Gaussian clusters as the smallest geometric support unit, the category probabilities within the cluster are weighted and aggregated according to data quality weights to generate fine-grained atomic semantic instances. Then, multiple sub-instances are aggregated from bottom to top according to the semantic granularity level to generate a virtual aggregated parent instance that does not require independent continuous geometric support. Simultaneously, four types of directed instance relationships are established: composition, attachment, location, and occupancy. For composition relationships, directed acyclic verification and single-parent constraints are performed to avoid erroneous confidence inheritance caused by mixing different relationships. Multi-level geometric detail layer payloads and multi-level semantic detail layer metadata are constructed separately, establishing independent mapping relationships between them. Based on geometric support coverage and view projection pixel area, the geometric evidence threshold corresponding to each semantic level is calculated to quantitatively determine whether the current geometric precision is sufficient to support the corresponding semantic granularity. Finally, the spatial index, semantic inverted index, instance relationship graph index, and hierarchical mapping index are integrated and encapsulated to generate an immutable scene snapshot with a unique snapshot identifier, ensuring that all indexes, instance relationships, and payload data point to the same version state.

[0026] Step S103: After parsing the query request, generate a query plan that is bound to the scene snapshot. Under the geometric evidence constraint, obtain the query results at the geometric and semantic levels from the scene snapshot according to the query plan.

[0027] The system analyzes the spatial scope, target category, target semantic granularity, task type, and resource budget of user query requests. It selects a valid scene snapshot and generates a standardized query plan bound to that snapshot identifier, unifying spatial predicates, semantic conditions, and constraint rules. A candidate instance set is obtained by taking the intersection of the spatial and semantic indexes. The geometric level, semantic level, and return pattern combination of each candidate instance are enumerated. First, a hard filter is applied using geometric evidence thresholds to eliminate invalid combinations with insufficient geometric support. Then, priority is calculated based on task relevance, semantic confidence, visual importance, and resource consumption. Within the bandwidth, memory, and response time budget, the optimal geometric and semantic level combination is selected, and query results with corresponding precision are output. For scenarios with insufficient resources, a downgraded return is performed according to the rule of "prioritizing semantic granularity while reducing geometric precision, then reducing semantic granularity." The entire retrieval is performed based on the same scene snapshot to avoid cross-version data conflicts, result duplication, and omissions.

[0028] Optionally, the generation process of atomic Gaussian clusters includes: Step S201: For the geometric attributes and semantic data of the three-dimensional Gaussian primitives in the target three-dimensional Gaussian splash scene, after verification and normalization processing, generate multiple Gaussian record data containing calibration category probability, mass weight and bounding box.

[0029] The process acquires the geometric attributes and multi-source semantic data of all 3D Gaussian primitives in the target 3D Gaussian splash scene, performs geometric verification and semantic normalization, and generates standardized Gaussian record data. Specifically, it first verifies the uniqueness of the Gaussian primitive's identifier, the finiteness of its coordinates, the legality of its scale, and the positive semidefiniteness of its covariance. It then performs normalization correction on repairable geometric errors and marks invalid data that cannot be repaired as excluded from the clustering candidate set. Based on the Gaussian center coordinates and the covariance matrix, it calculates a bounding box with three standard deviations, which serves as the unified geometric range for subsequent spatial partitioning and adjacency search.

[0030] A unified category mapping is performed on the multi-source semantic outputs from segmentation models, detection models, or manual annotations, mapping heterogeneous labels to the standard ontology category system. When a unique mapping is not possible, multiple candidate categories and their corresponding probabilities are retained. The category scores are probabilistically calibrated using the temperature parameters corresponding to each source to reduce the inherent scoring bias of different models. The quality weight of each piece of evidence is calculated by combining the reliability of the source, the quality of the viewpoint, and the geometric reliability. The validity of each field is marked item by item through the effective feature mask, a unique evidence identifier is assigned to each observation, and global evidence deduplication is performed to ensure that the same evidence is only accumulated once in subsequent hierarchical aggregation. Finally, standardized Gaussian record data containing calibrated category probabilities, quality weights, effective feature labels, and bounding boxes with three times the standard deviation is generated.

[0031] Step S202: Recursively divide the target 3D Gaussian splash scene into spatial nodes according to the number of Gaussians and spatial size within the master-owned leaf node. The dividing surface is selected at the mass-weighted median of the longest axis of the node, and the leaf node where the center of each Gaussian record data is located is taken as the unique master-owned leaf node.

[0032] Based on the spatial distribution of standardized Gaussian record data, a recursively subdividable spatial hierarchy tree is constructed, and a unique master-owned leaf node is determined for each Gaussian record. Specifically, the root node is constructed using the extreme range of the Gaussian bounding box of the entire scene, simultaneously recording the scene identifier, coordinate system, and data version. When any of the Gaussian quantity, spatial size, or estimated visual error under the reference view exceeds the corresponding threshold, the longest axis of the node is selected, and the dividing plane is determined according to the quality-weighted median of the Gaussian center position. The node is recursively divided into two non-empty child nodes, and the child node identifier is generated deterministically according to the spatial position, ensuring a consistent spatial structure when the same input is repeatedly executed. The leaf node where the center coordinates of the Gaussian record are located is taken as its unique master-owned node, and the complete Gaussian data is written into the ownership set of the master-owned node. For Gaussians whose bounding boxes cross the node boundaries, only lightweight references containing Gaussian identifiers, the owner node, and the intersection range are registered in adjacent nodes, without copying the complete geometric and semantic data, thus eliminating cross-block duplicate ownership from the data structure perspective.

[0033] After the construction is completed, an integrity check is performed to confirm that all Gaussian nodes are unique and have no overlap, and that cross-node references can be traced back to the unique owner node, thus providing a structural foundation for globally unique clustering.

[0034] Step S203: Within each primary leaf node, the connection weights of candidate pairs are calculated through nearest neighbor search. Then, the connection weights are used to cluster and merge all Gaussian record data according to nodes and boundaries to obtain the atomic Gaussian clusters corresponding to each Gaussian record data.

[0035] A two-stage strategy of intra-node local clustering and cross-boundary global merging is adopted to perform clustering and merging on all Gaussian record data, generating globally unique and non-overlapping atomic Gaussian clusters. Specifically, in the first stage, within each leaf node, candidate Gaussian pairs with spatial distances within a threshold range are retrieved using a nearest neighbor index. Connection weights are calculated based on four dimensions: central distance, bounding box overlap, normal consistency, and semantic similarity. When features are missing, the weights are renormalized according to the valid features to avoid missing items being misclassified as zero. After candidate edges are stably sorted in descending order of weight and lexicographical order of Gaussian identifiers, temporary intra-node clusters are generated by combining candidate pairs that satisfy the intra-node thresholds using a disjoint-set search. In the second stage, only candidate Gaussian pairs at the boundaries of adjacent nodes are processed. Connection features and weights are recalculated, and candidate edges that satisfy the cross-boundary thresholds are processed sequentially according to a stable sorting. The corresponding temporary clusters are generated by combining them using a global disjoint-set search. The order of all boundary processing is stably generated based on Gaussian identifiers and is not affected by the parallel scheduling order, ensuring the repeatability of clustering results.

[0036] After global merging is completed, the Gaussian identifiers within each connected component are sorted and a content hash is generated as a cluster identifier. It is verified that each Gaussian belongs to only one cluster and all valid Gaussians have been clustered, finally obtaining atomic Gaussian clusters with unique geometric ownership and no overlap.

[0037] Optionally, the process of generating the semantic instance graph includes: Step S301: For the category probabilities of Gaussian record data in the atomic Gaussian cluster, perform weighted aggregation according to preset quality weights to generate corresponding fine-grained semantic instances; wherein, the fine-grained semantic instances adopt a mechanism of separating stable identity identifiers and immutable revision identifiers, and modifying the semantics only generates new revisions without changing the identity.

[0038] Using globally unique and non-overlapping atomic Gaussian clusters as the smallest geometric support unit, the calibration category probabilities of all Gaussian records within the cluster are weighted and normalized according to their corresponding quality weights to obtain the category probability distribution at the atomic cluster level. During the aggregation process, global deduplication is performed based on the unique evidence identifier, and the same evidence is only accumulated once when propagating along multiple Gaussians to avoid the repeated amplification of a single observation.

[0039] When the highest category confidence reaches a preset threshold, a corresponding fine-grained semantic instance is generated. When the difference in confidence between the top two categories is less than the ambiguity threshold, multi-category candidates and conflict state markers are retained, and disambiguation is completed after upper-level relation constraints or human evidence intervention. An identity mechanism that separates stable identity identifiers and immutable revision identifiers is adopted: stable identity identifiers correspond to the same semantic object in the real world and remain unchanged across semantic revisions; immutable revision identifiers record the complete state of a single semantic description of the object, and when semantic modifications, relation adjustments, or geometric support changes occur, only a new revision identifier is generated, without overwriting historical versions or changing the object's stable identity.

[0040] Step S302: After aggregating multiple sub-instances in the fine-grained semantic instance according to the execution order corresponding to the semantic level, a virtual aggregated parent instance without independent geometric support is generated.

[0041] Following a semantic granularity hierarchy from fine to coarse, bottom-up aggregation processing is performed on component-level, part-level, and object-level semantic instances sequentially to generate virtual aggregated parent instances without independent continuous geometric support. For each target parent category, the set of allowed subcategories is obtained from the preset ontology composition relationship template. Only sub-instances within the same spatial node and its neighborhood are selected as aggregation candidates. When the bounding box distance between two sub-instances is not greater than the spatial threshold of the corresponding parent category, the boundary continuity support is not lower than the preset threshold, and the diagonal of the merged bounding box does not exceed the upper limit of the parent category size, candidate aggregation edges are established between the sub-instances. After the candidate aggregation edges are sorted in descending order of comprehensive score and stable lexicographical order of instance identifiers, candidate components are formed using a disjoint-set data structure algorithm. The upper limit of size is re-verified before each merging to avoid forming incorrect parent objects through chain aggregation of distant objects.

[0042] Then, a virtual aggregate parent instance is created for each final candidate component. Its geometric support set is the deduplicated union of the atomic Gaussian clusters corresponding to all child instances. There is no need for an independent parent Gaussian cluster, which is suitable for scenarios where high-level semantic objects are expressed by multiple discontinuous geometric clusters.

[0043] Step S303: Generate a semantic instance graph corresponding to the atomic Gaussian cluster based on one or more directed relationships among composition, attachment, location, and occupancy between each sub-instance domain and the virtual aggregate parent instance.

[0044] Based on the association between child instances and virtual aggregated parent instances, a semantic instance graph containing various directed relation types is constructed, with different constraints and propagation rules applied to different relation types. Specifically, composition relations express whole-part subordination relationships between instances, retaining only the highest-scoring parent instance on the same semantic main path and performing directed acyclic detection to prevent the formation of circular subordination relationships; attachment relations express installation and fitting relationships between instances, allowing multiple attached parents and not implementing category confidence inheritance; positional relations express spatial containment relationships between instances, allowing multi-level nested positional links; and occupancy relations express the space occupancy relationship of temporary objects, not recognizing occupier objects as subclasses of occupied space.

[0045] The confidence scores for each type of relationship, the relationship confidence scores, and the parent aggregation confidence scores are stored independently to avoid confusing the certainty of category identification with the certainty of relationship establishment; all relationship edges are associated with evidence references, and the parent aggregation is calculated by deduplicating the evidence identifiers to ensure the accuracy of confidence propagation.

[0046] Optionally, the process of acquiring a scene snapshot includes: Step S401: Using the intersection relationships, semantic content, instance graphs and hierarchical mappings corresponding to the geometric level load and semantic level data, construct spatial indexes that support range queries and view frustum queries, semantic inverted indexes that point from categories to instances, instance graph indexes that store type relationships, and mapping indexes that represent detail levels.

[0047] A multi-level geometric hierarchy of payloads, from coarse to fine, is constructed based on atomic Gaussian clusters. Each payload level records the support weight mapping of the corresponding original Gaussian. Gaussian data of different precisions can be generated through error-constrained simplification and hierarchical sampling. A multi-level semantic hierarchy of metadata with corresponding granularity is constructed based on semantic instance graphs, storing only lightweight category, relationship, confidence, and spatial range information. Four types of indexes are built upon this foundation: First, a spatial index, using the bounding box of each semantic instance as the retrieval unit, supports spatial range intersection queries, view frustum clipping queries, and proximity searches for fast spatial filtering of candidates; second, a semantic inverted index, establishing an inverted mapping from standard ontology category nodes to semantic instance identifiers, supporting fast instance retrieval by category, and configurable for recursive subclass expansion; third, an instance graph index, storing directed adjacency lists categorized by relationship type, supporting path traversal and hierarchical expansion for different relationships such as composition, attachment, location, and occupancy; and fourth, a hierarchical mapping index, using instance identifiers and target semantic levels as retrieval keys to locate corresponding available geometric hierarchy payloads, metadata references, and geometric evidence thresholds, achieving independent addressing at the geometric and semantic levels.

[0048] Step S402: After pointing the spatial index, semantic inverted index, instance graph index and mapping index to the same snapshot identifier, obtain the instance revision set and load list corresponding to the atomic Gaussian cluster using geometric level load and semantic level data.

[0049] Perform a snapshot consistency pre-verification to confirm that the spatial index, semantic inverted index, instance graph index, and hierarchical mapping index are all associated with the same snapshot identifier to be released, and that all instances, relationships, and payloads pointed to by all indexes originate from the same scenario version. Simultaneously verify the directed acyclicity of the constituent relationships, the global uniqueness of Gaussian attribution, and the deduplication completeness of evidence aggregation; if any verification fails, the release process is not initiated. After successful verification, compile the instance revision set corresponding to all atomic Gaussian clusters, fully including the stable identity identifier, immutable revision identifier, category attribute, relationship path, confidence level, and spatial range of each semantic instance. Simultaneously compile the geometric payload list corresponding to each precision and the semantic metadata list for each granularity, establishing a supporting correspondence between instances and payloads to ensure that the index can accurately trace back to the underlying geometric and semantic data.

[0050] Step S403: Construct a scene snapshot of the atomic Gaussian cluster using the same snapshot identifier, instance revision set, and load list.

[0051] The root nodes, complete instance revision sets, geometric payload lists, semantic metadata lists, hierarchical mapping rules, and geometric evidence threshold parameters corresponding to each semantic level of the four types of indexes are uniformly encapsulated to generate an immutable scene snapshot with a globally unique snapshot identifier. An atomic release mechanism is adopted, exposing the snapshot identifier and access entry point to the outside world at once after all data assembly and verification are completed, prohibiting the premature release of incomplete incomplete indexes; after the snapshot is released, its internal data remains unmodifiable, and all query, pagination, and continued retrieval operations are executed based on this fixed snapshot state, ensuring the version consistency of spatial data, semantic data, instance relationships, and payload references in the entire retrieval chain, and fundamentally avoiding result duplication, omissions, and conflicts caused by cross-version data mixing.

[0052] Optional, the process of constructing geometric evidence constraints includes: Step S501: Obtain the geometric detail level load corresponding to the semantic instance graph based on the semantic level data, and record the original support weight corresponding to the semantic instance graph through the geometric detail level load.

[0053] For each semantic instance, a multi-level geometric detail load is generated, from coarse to fine. Each load level is generated through clustering representation, error-constrained simplification, or stratified sampling. Each simplified Gaussian data point is associated with a support weight mapping field, recording the original Gaussian set it represents and its corresponding weight percentage. Based on the total support weight of the full-precision geometric load, the support weight coverage of the current geometric level load is calculated, which is the ratio of the sum of the original support weights represented by the current level to the total full support weight. This is used to quantify the geometric information completeness of the current geometric level, providing a foundation for subsequent evidence quantification.

[0054] Step S502: Calculate the geometric evidence score corresponding to the original support weight using the projected pixel area of ​​the semantic instance graph in the current view.

[0055] By combining the query viewpoint and view frustum parameters, the projected area of ​​the spatial bounding box of the semantic instance under the current viewpoint is calculated and converted into the projected pixel area on the display plane. The projected pixel area is then normalized and truncated with the minimum recognition projected area threshold corresponding to the target semantic level to obtain the viewpoint adaptation coefficient. The aforementioned geometric support weight coverage rate is multiplied by the viewpoint adaptation coefficient to obtain the geometric evidence score. This score reflects both the accuracy and completeness of the geometry itself and the degree of recognition under the current viewpoint, enabling a quantifiable assessment of geometric support capabilities.

[0056] Step S503: Construct geometric evidence constraints for each semantic level in the semantic instance graph using geometric evidence scores.

[0057] Each semantic level is assigned a corresponding geometric evidence threshold, forming hierarchical and differentiated geometric evidence constraints. When the geometric evidence score is not lower than the threshold for the corresponding level, the current geometric level is deemed sufficient to support the query and use of that semantic level; when it is lower than the threshold, the geometric support is deemed insufficient, and the direct output of the complete geometric result for the corresponding semantic level is prohibited. Furthermore, the threshold can be configured differently based on the task type. High thresholds are used for security-sensitive tasks such as navigation, collision detection, and precise measurement, and pure semantic return mode is disabled. Ordinary browsing and retrieval tasks can use an adaptive threshold, supporting downgraded return of pure semantic metadata, thus achieving scenario-based adaptation of the constraints.

[0058] Optional, the query plan generation process includes: Step S601: After obtaining the query request, convert the bounding box parameter, radius parameter and view frustum parameter in the query request to the scene coordinate system corresponding to the scene snapshot.

[0059] In response to a query request, the system first selects a scene snapshot identifier that is still valid from the available snapshot directory as the sole data version benchmark for this query. All spatial parameters included in the query request, such as bounding box, search radius, view frustum, and probe ray, are uniformly converted to the standard scene coordinate system corresponding to that scene snapshot. The system simultaneously records the corresponding coordinate system number, spatial tolerance threshold, and whether a parent context needs to be returned, thus preventing retrieval bias caused by direct comparison of spatial ranges from different coordinate systems.

[0060] Step S602: Construct spatial transformation conditions, semantic candidate conditions, and task constraints corresponding to the query request based on the scene coordinate system.

[0061] Based on the transformed spatial parameters, standardized spatial predicates are generated, supporting various spatial retrieval modes such as range intersection, radial proximity, view frustum clipping, and ray penetration. The target semantic description in the query request is converted into a standard category node set through ontology synonym mapping, and the extension identifier of whether to recursively expand subcategories is configured. When the ambiguity exceeds a preset threshold, an upper-layer application confirmation mechanism is triggered. According to the task type corresponding to the query, the target semantic level, minimum geometric evidence threshold, allowed return mode, risk weight, and required return fields are configured differently. Resource budgets such as time, network bandwidth, video memory usage, and number of objects are uniformly converted into accumulative quantitative units to form complete query constraints.

[0062] Step S603: Generate a unique query signature using the snapshot identifier, spatial transformation conditions, semantic candidate conditions, and task constraints of the scene snapshot, and generate a query plan for the scene snapshot based on the query signature.

[0063] The scene snapshot identifier, standardized spatial predicate, target category node set, target semantic level, task type, sorting rules, return fields and budget constraints are serialized in a preset deterministic order to calculate and generate a globally unique query signature. Based on all the above elements, a standardized query plan is generated. Subsequent spatial retrieval, semantic recall, level selection, result return and pagination continuation query operations are all bound to the scene snapshot identifier and query signature for execution, ensuring that the same query reads the same scene version, follows the same constraint rules and outputs the same results throughout the entire process.

[0064] Optionally, the process of obtaining query results includes: Step S701: Obtain the spatial index and semantic index corresponding to the query plan from the scene snapshot using geometric level load and semantic level data, respectively.

[0065] Based on the scene snapshot identifier bound to the query plan, the index root node of the corresponding immutable scene snapshot is located, and the spatial index and semantic inverted index are retrieved respectively. The spatial index loads the instance bounding box and spatial retrieval structure to respond to the spatial predicate conditions in the query plan. The semantic inverted index loads the mapping relationship from standard category nodes to semantic instances to respond to the target category retrieval and subclass expansion requirements. The index data is read based on the same scene snapshot throughout the process to avoid result deviations caused by mixing cross-version indexes.

[0066] Step S702: Obtain a candidate instance set by using the intersection of the spatial index and the semantic index, and obtain the combination result between the geometric level, semantic level and return pattern of each instance in the candidate instance set.

[0067] Based on the spatial predicate in the query plan, the spatial index is traversed to obtain a first set of instances that meet the conditions of spatial range, view frustum, or proximity. Based on the target category node set and subclass extended identifier, the semantic inverted index is traversed to retrieve a second set of instances corresponding to the category. The intersection of the first and second instance sets is taken to form a candidate instance set. For each semantic instance in the candidate set, multiple feasible combinations of geometric detail level, semantic detail level, and return mode are independently enumerated. The return mode includes at least two types: joint geometric and semantic return, and pure semantic metadata return. Simultaneously, attribute parameters such as geometric payload volume, memory usage, spatial error, and minimum geometric evidence threshold corresponding to each combination are read to provide quantitative basis for subsequent filtering and sorting.

[0068] Step S703: Generate the query results corresponding to the query request by combining the results that satisfy the geometric evidence constraints.

[0069] Using geometric evidence constraints as a hard prerequisite, a geometric evidence score is calculated for each combination, and invalid combinations with scores below the threshold corresponding to the target semantic level are eliminated. For safety-sensitive tasks such as navigation and collision detection, the pure semantic return mode is directly disabled to ensure the reliability of geometric support. For feasible combinations that pass the threshold check, priority weights are calculated by comprehensively considering task relevance, semantic confidence, perspective importance, and resource consumption. The combinations are then stably sorted from high to low priority and selected sequentially within the time, bandwidth, and memory budget of the query plan. When the target combination exceeds the budget, an adaptation is performed following a degradation strategy of "prioritizing maintaining the semantic level, reducing the geometric level, and then reducing the semantic level according to rules," and the corresponding degradation reason code is recorded. Finally, a query result set that meets the constraints is generated.

[0070] Optionally, the method further includes: Step S801: Obtain the data space occupied by the query results in real time, and use the data space occupied to generate the return batch corresponding to the query request.

[0071] Adhering to the stable sorting rules of query results (descending priority, ascending semantic instance identifier, ascending geometric level, and descending semantic level), query result entries are assembled item by item. Each result carries an instance stability identifier, revision identifier, category attribute, relationship path, corresponding geometric / semantic level, geometric evidence score, spatial error, and status reason code. The total data bytes, instance count, and expected response latency of the current batch are accumulated in real time. When any indicator reaches the batch limit set in the query plan, the current batch is truncated. The sorting key corresponding to the last result in the current batch is recorded as the pagination anchor, replacing traditional array subscripts to avoid pagination misalignment caused by data insertion or reordering. Geometric payloads can be directly embedded or returned as payload references with verification identifiers, which are loaded by the client on demand, further reducing invalid transmission.

[0072] Step S802: Obtain the query signature of the query request, and generate a token verification result of the query signature using the query identifier and snapshot identifier corresponding to the returned batch.

[0073] Based on the current batch's pagination anchor point, remaining resource budget, and validity period, combined with the query identifier, scene snapshot identifier, and original query signature in the query plan, a further refinement token is generated. A message authentication code is calculated from the token's core fields using the server-side key as a verification value to prevent client-side tampering with the budget, snapshot version, or pagination position. Upon receiving a continued query request, the token's legality and validity period are first verified. Then, the query signature carried on the token is compared with the original query signature to confirm that the query conditions have not been tampered with. Simultaneously, it is verified that the corresponding scene snapshot is still in a valid and readable state, generating a token verification result.

[0074] Step S803: Update the query results using the token verification result.

[0075] When token verification passes, the retrieval logic and constraints of the original query plan are reused, and the search continues from the pagination anchor point of the token record. The instance stability identifier plus the revision identifier is used as the global deduplication key to avoid duplicate results caused by network retries. When verification reveals that the scene snapshot has expired, a snapshot expiration status code is returned, forcing the client to regenerate the query plan based on the latest scene snapshot. This prevents the direct use of old pagination anchor points in new version scenarios, avoiding duplicate and omission results caused by cross-version data reading. Simultaneously, a status reason code is attached to each downgraded or skipped instance to facilitate explanation and decision-making by upper-layer applications.

[0076] refer to Figure 2 Another data query method for 3D Gaussian splash scenes, as shown, includes the following steps: Step S100: Obtain and normalize the three-dimensional Gaussian semantic evidence.

[0077] The process involves acquiring the geometric attributes and multi-source semantic data of all Gaussian primitives in a 3D Gaussian splash scene, performing full-dimensional normalization preprocessing, and generating standardized Gaussian records. Specifically, the process first performs geometric verification on the uniqueness of Gaussian identifiers, the finiteness of coordinates, and the positive semidefiniteness of covariance, correcting normalizable geometric errors, eliminating invalid data, and calculating a bounding box with three standard deviations based on the center coordinates and covariance matrix as a unified spatial metric for subsequent spatial partitioning and adjacency calculation. Heterogeneous semantic labels from segmentation models, detection models, and manually labeled data are mapped to a standard ontology category system, and the category scores are probabilistically calibrated using the temperature parameters corresponding to each source to eliminate inherent scoring biases from different models. The quality weight of each piece of semantic evidence is calculated by comprehensively considering source reliability, viewpoint quality, and geometric reliability. The validity of each field is marked item by item using a feature effective mask, a unique identifier is assigned to each piece of evidence, and global deduplication is performed to ensure that the same evidence is only accumulated once in subsequent hierarchical aggregation, thus unifying the data caliber from the source.

[0078] Step S200: Construct spatial hierarchy and determine unique primary affiliation.

[0079] Based on the spatial distribution characteristics of standardized Gaussian records, a recursively subdividable spatial hierarchy tree is constructed, and a unique master-owned node is established for each Gaussian. The root node is constructed using a full-scene bounding box. When the number of Gaussians within a node, its spatial span, or the estimated visual error from the reference viewpoint exceeds a preset threshold, the longest axis of the node is selected, and a dividing plane is determined based on the quality-weighted median position of the Gaussian center, recursively dividing the node into two non-empty child nodes. Node identifiers are deterministically generated based on spatial location, ensuring complete consistency of spatial structure when the same input is repeatedly executed. The leaf node containing the Gaussian center is used as its unique master-owned node. Complete Gaussian data is only written to the ownership set of the master-owned node. For Gaussians whose bounding boxes cross node boundaries, only lightweight references containing identifiers, the master-owned node, and the intersection range are registered in adjacent nodes; complete geometric and semantic data are not copied. After construction, integrity verification is performed to confirm that all Gaussian ownership is unique and non-overlapping, and cross-node references can be traced back to the unique master-owned node, eliminating the root cause of duplicate storage across blocks from a data structure perspective.

[0080] Step S300: Generate a globally unique Gaussian cluster of atoms through a two-stage merging process.

[0081] A two-stage strategy of intra-node local clustering and cross-boundary global merging is adopted to generate globally unique and non-overlapping atomic Gaussian clusters. In the first stage, within each leaf node, candidate Gaussian pairs with spatial distances within a threshold range are retrieved using a nearest neighbor index. Connection weights are calculated based on four dimensions: center distance, bounding box overlap, normal consistency, and semantic similarity. When features are missing, weights are renormalized according to valid features to avoid misclassifying missing items as zero. After candidate edges are stably sorted by weight in descending order and Gaussian identifier lexicographical order, they are merged into intra-node temporary clusters using a disjoint-set lookup and candidate pairs satisfying the intra-node threshold. In the second stage, only candidate Gaussian pairs at adjacent node boundaries are processed. Connection features and weights are recalculated, and candidate edges satisfying the cross-boundary threshold are processed sequentially according to a stable sorting. These are then merged into corresponding temporary clusters using a global disjoint-set lookup. The order of all boundary processing is based on the stable generation of Gaussian identifiers and is unaffected by the parallel scheduling order. After global merging is completed, the Gaussian identifiers of each connected component are sorted and a content hash is generated as a cluster identifier. Each Gaussian is verified to belong to only one cluster, and finally, atomic Gaussian clusters with unique geometric ownership and no overlap are obtained, which serve as the smallest geometric support unit for subsequent semantic organization.

[0082] Step S400: Generate a typed semantic instance graph from the bottom up.

[0083] Based on atomic Gaussian clusters as the geometric foundation, a semantic instance graph with stable identities and distinguishable relationship types is constructed from the bottom up. First, the category probabilities within atomic clusters are aggregated using quality weights to generate fine-grained atomic semantic instances. A mechanism separating stable identity identifiers from immutable revision identifiers is adopted. The stable identity identifier corresponds to the same object in reality and remains unchanged across semantic revisions; semantic modifications only generate new revision identifiers without overwriting historical versions. Following a fine-to-coarse semantic granularity order, aggregated child instances generate virtual aggregated parent instances without independent continuous geometric support. The geometric support of the parent instance is the deduplicated union of the atomic clusters corresponding to all child instances, adapting to scenarios where high-level objects are expressed by multiple discontinuous geometric clusters. Four types of directed relationships are constructed: composition, attachment, position, and occupancy. Different constraints are applied to different relationships: composition relationships retain only a unique parent level on the same main path and perform directed acyclic detection; attachment relationships allow multiple parents and do not pass category confidence; position relationships allow multi-level spatial nesting; occupancy relationships do not recognize the occupied object as a spatial subclass. Each confidence level is stored independently, and the parent aggregation is performed by deduplicating the evidence identifier to avoid incorrect confidence level inheritance and relationship conflicts.

[0084] Step S500: Establish dual LOD mapping, dual indexes, and immutable scene snapshots.

[0085] Geometric detail level payloads and semantic detail level metadata are constructed separately, establishing independent dual-LOD mappings and encapsulating them as immutable scene snapshots. For each semantic instance, multi-level geometric LOD payloads are generated from coarse to fine. Each level of payload is simplified through error constraints and generated through hierarchical sampling, recording the corresponding original Gaussian support weight mapping. Simultaneously, multi-level semantic LOD metadata is generated, storing only lightweight category, relationship, confidence, and spatial range information. A dual-LOD mapping relationship is established, and geometric evidence scores are calculated based on geometric support weight coverage and view projection pixel area. Corresponding evidence thresholds are set for each semantic level to quantify whether the current geometric precision is sufficient to support the corresponding semantic granularity. Four types of indexes are constructed: spatial index, semantic inverted index, instance graph index, and LOD mapping index, supporting spatial retrieval, category recall, relationship traversal, and hierarchical addressing, respectively. The four types of indexes, instance revision sets, payload lists, and mapping rules are uniformly encapsulated, and after performing consistency checks, are published once as an immutable scene snapshot. All data remains unmodifiable after publication, ensuring that the entire chain reads the same version state.

[0086] Step S600: Parse the query and generate a snapshot query plan.

[0087] User queries are parsed into standardized query plans bound to fixed scene snapshots, ensuring repeatable and consistent retrieval processes. Upon query initiation, a valid scene snapshot identifier is selected and written into the query plan; all subsequent operations are executed based on this snapshot. Spatial parameters such as bounding boxes, view frustums, and radii are uniformly converted to the scene coordinate system, generating standardized spatial predicates. Target semantic descriptions are mapped to a set of standard category nodes, and subclass extension rules are configured. Target semantic levels, minimum evidence thresholds, allowed return modes, and risk weights are configured differently based on task type, and resource budgets such as time, bandwidth, and video memory are converted into accumulative quantitative constraints. Snapshot identifiers, query conditions, and constraint rules are serialized in a defined order to generate a unique query signature; subsequent paginated queries must use the same signature to ensure that query conditions are not tampered with.

[0088] Step S700: Select the geometric level and the semantic level respectively under the constraints of evidence and budget.

[0089] Under the rigid constraint of the geometric evidence threshold, the optimal combination of geometric and semantic levels is independently selected for each candidate instance based on resource budget. A set of candidate instances is obtained by taking the intersection of spatial and semantic indices, and feasible combinations of geometric, semantic, and return modes for each instance are enumerated. First, the geometric evidence threshold is used for pre-filtering, eliminating invalid combinations with insufficient geometric support; pure semantic return modes are directly disabled for safety-sensitive tasks such as navigation and collision detection to ensure result reliability. Feasible combinations are calculated based on task relevance, semantic confidence, viewpoint importance, and resource consumption priority, and are stably sorted from high to low priority, with results selected sequentially within the budget. When the budget is exceeded, a strategy of "prioritizing maintaining the semantic level, reducing the geometric level, and then reducing the semantic level according to rules" is followed to perform degradation, and the reason for degradation is recorded to achieve a dynamic balance between accuracy and resources.

[0090] Step S800: Return the results and continue refining the tokens within the same snapshot.

[0091] The results are batched according to a stable sorting order, and a further refinement token bound to the same snapshot is generated to ensure that the pagination process is not duplicated, omitted, or cross-version. Results are assembled item by item using a stable sorting order based on priority and instance identifier. The batch is truncated when the cumulative data volume, number of instances, or response latency reaches the batch limit. The sorting key of the last result is recorded as the pagination anchor, replacing traditional array indices to avoid pagination misalignment caused by data reordering. A further refinement token containing query identifier, snapshot identifier, query signature, pagination anchor, remaining budget, and verification value is generated, and a message authentication code prevents client-side parameter tampering. Upon receiving a continuation query request, the token's validity and snapshot validity are verified first. If the verification passes, the retrieval continues from the pagination anchor. If the snapshot has expired, the expired state is returned, and a new query is forced based on the new snapshot. The use of old pagination anchors across versions is prohibited, thus preventing result duplication and omissions caused by cross-version mixed readings.

[0092] The data query method described above for 3D Gaussian splash scenarios demonstrates that this method establishes spatial hierarchy and unique master affiliation for 3D Gaussian primitives. Within spatial nodes, it performs global merging based on cross-node boundary edges to obtain non-overlapping atomic Gaussian clusters. Then, based on these atomic Gaussian clusters, it generates semantic instances with stable identities, thereby generating an immutable scene snapshot that can aggregate multiple atomic Gaussian clusters. Geometric evidence constraints corresponding to each semantic level are established within this scene snapshot. During querying, the method returns corresponding geometric and semantic, semantic-only, or downgraded results based on geometric evidence and resource budget. This method reduces cross-block duplication, avoids erroneous semantic inheritance, and reduces irrelevant geometric transfers and GPU memory usage required for fine-grained semantic queries.

[0093] Corresponding to the above-described data query method embodiment for a 3D Gaussian splash scene, this embodiment of the invention also provides a data query system for a 3D Gaussian splash scene, such as... Figure 3 As shown, the system includes: Atomic Gaussian Cluster Generation Module 100: Used to construct a spatial hierarchy tree using scene location data in the target 3D Gaussian splash scene, determine a unique master-belonging leaf node for each Gaussian record data in the target 3D Gaussian splash scene based on the spatial hierarchy tree, and use the master-belonging leaf node to perform clustering and merging processing on all Gaussian record data to generate atomic Gaussian clusters; Constraint generation module 200: It is used to construct geometric level loads and semantic level data respectively based on the semantic instance graph generated by atomic Gaussian clusters, integrate the intersection relationship, semantic content, instance graph and level mapping corresponding to the geometric level load and semantic level data into scene snapshot, and construct the geometric evidence constraints corresponding to each semantic level. Query result acquisition module 300: used to generate a query plan that binds to the scene snapshot after parsing the query request, and obtain the geometric and semantic query results from the scene snapshot according to the query plan under the geometric evidence constraint.

[0094] The data query system for 3D Gaussian splash scenarios described above demonstrates that it can establish spatial hierarchies and unique primary affiliations for 3D Gaussian primitives. Within spatial nodes, it performs global merging based on cross-node boundary edges to obtain non-overlapping atomic Gaussian clusters. Then, based on these atomic Gaussian clusters, it generates semantic instances with stable identities, thereby generating an immutable scene snapshot that can aggregate multiple atomic Gaussian clusters. Geometric evidence constraints corresponding to each semantic level are established within this scene snapshot. During queries, the system returns corresponding geometric and semantic, semantic-only, or downgraded results based on geometric evidence and resource budget. This system reduces cross-block duplication, avoids erroneous semantic inheritance, and lowers the irrelevant geometric transfers and GPU memory usage required for fine-grained semantic queries.

[0095] The data query system for three-dimensional Gaussian splash scenes provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned data query method embodiment for three-dimensional Gaussian splash scenes. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned data query method embodiment for three-dimensional Gaussian splash scenes.

[0096] This embodiment also provides an electronic device, the structural schematic diagram of which is shown below. Figure 4 As shown, the device includes a processor 101 and a memory 102; wherein, the memory 102 is used to store one or more computer instructions, which are executed by the processor to implement the steps of the above-described data query method for a three-dimensional Gaussian splash scene.

[0097] Figure 4 The electronic device shown also includes a bus 103 and a communication interface 104, with the processor 101, communication interface 104 and memory 102 connected via the bus 103.

[0098] The memory 102 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. The bus 103 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0099] The communication interface 104 is used to connect to at least one user terminal and other network units through a network interface, and to send encapsulated IPv4 packets or IPv4 packets to the user terminal through the network interface.

[0100] Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. The processor 101 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 102. The processor 101 reads the information in memory 102 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0101] This invention also provides a storage medium storing a computer program, which, when run by a processor, executes the steps of the data query method for a three-dimensional Gaussian splash scene described in the foregoing embodiments.

[0102] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, devices, and methods can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0103] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0104] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0105] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A data query method for a three-dimensional Gaussian splash scene, characterized in that, The method includes: A spatial hierarchy tree is constructed using scene location data in the target 3D Gaussian splash scene. Based on the spatial hierarchy tree, a unique master-belonging leaf node is determined for each Gaussian record data in the target 3D Gaussian splash scene. The master-belonging leaf node is then used to cluster and merge all Gaussian record data to generate an atomic Gaussian cluster. Based on the semantic instance graph generated by the atomic Gaussian cluster, geometric level payloads and semantic level data are constructed respectively. The intersection relationship, semantic content, instance graph and level mapping corresponding to the geometric level payload and the semantic level data are integrated into a scene snapshot, and geometric evidence constraints corresponding to each semantic level are constructed. After parsing the query request, a query plan is generated that is bound to the scene snapshot. Under the geometric evidence constraint, the query plan is used to obtain the geometric and semantic query results from the scene snapshot.

2. The data query method for a three-dimensional Gaussian splash scene according to claim 1, characterized in that, The generation process of the atomic Gaussian cluster includes: For the geometric properties and semantic data of the three-dimensional Gaussian primitives in the target three-dimensional Gaussian splash scene, after verification and normalization processing, multiple Gaussian record data containing calibration category probability, mass weight and bounding box are generated. The target 3D Gaussian splash scene is recursively divided into spatial nodes according to the number of Gaussians and spatial size in the master-owned leaf node. The dividing surface is selected at the mass-weighted median of the longest axis of the node, and the leaf node where the center of each Gaussian record data is located is the unique master-owned leaf node. Within each of the primary leaf nodes, the connection weights of candidate pairs are calculated through nearest neighbor search. Then, the connection weights are used to cluster and merge all Gaussian record data according to nodes and boundaries to obtain the atomic Gaussian clusters corresponding to each Gaussian record data.

3. The data query method for a three-dimensional Gaussian splash scene according to claim 1, characterized in that, The process of generating the semantic instance graph includes: For the category probabilities of Gaussian record data within the atomic Gaussian cluster, a weighted aggregation is performed according to a preset quality weight to generate corresponding fine-grained semantic instances; wherein, the fine-grained semantic instances adopt a mechanism that separates stable identity identifiers from immutable revision identifiers, and modifying the semantics only generates new revisions without changing the identity; After aggregating multiple sub-instances in the fine-grained semantic instance according to the execution order corresponding to the semantic level, a virtual aggregated parent instance without independent geometric support is generated. Generate a semantic instance graph corresponding to the atomic Gaussian cluster based on one or more directed relationships among composition, attachment, position, and occupancy between each sub-instance domain and the virtual aggregate parent instance.

4. The data query method for a three-dimensional Gaussian splash scene according to claim 1, characterized in that, The process of obtaining the scene snapshot includes: Using the intersection relationships, semantic content, instance graphs and hierarchical mappings corresponding to the geometric level load and the semantic level data, a spatial index supporting range queries and view frustum queries, a semantic inverted index pointing from categories to instances, an instance graph index storing type relationships, and a mapping index representing the detail level are constructed respectively. After pointing the spatial index, the semantic inverted index, the instance graph index, and the mapping index to the same snapshot identifier, the instance revision set and load list corresponding to the atomic Gaussian cluster are obtained using the geometric level load and the semantic level data. A scene snapshot of the atomic Gaussian cluster is constructed using the same snapshot identifier, the instance revision set, and the load list.

5. The data query method for a three-dimensional Gaussian splash scene according to claim 1, characterized in that, The process of constructing the geometric evidence constraints includes: Based on the semantic level data, the geometric detail level load corresponding to the semantic instance graph is obtained, and the original support weight corresponding to the semantic instance graph is recorded through the geometric detail level load; The geometric evidence score corresponding to the original support weight is calculated using the projected pixel area of ​​the semantic instance graph in the current view. Geometric evidence constraints for each semantic level in the semantic instance graph are constructed using the geometric evidence scores.

6. The data query method for a three-dimensional Gaussian splash scene according to claim 1, characterized in that, The process of generating the query plan includes: After obtaining the query request, the bounding box parameters, radius parameters, and view frustum parameters in the query request are converted to the scene coordinate system corresponding to the scene snapshot; Based on the scene coordinate system, construct the spatial transformation conditions, semantic candidate conditions, and task constraints corresponding to the query request; A unique query signature is generated using the snapshot identifier of the scene snapshot, the spatial transformation conditions, the semantic candidate conditions, and the task constraints, and a query plan for the scene snapshot is generated based on the query signature.

7. The data query method for a three-dimensional Gaussian splash scene according to claim 1, characterized in that, The process of obtaining the query results includes: The spatial index and semantic index corresponding to the query plan are obtained from the scene snapshot using the geometric level load and the semantic level data, respectively. A candidate instance set is obtained by using the intersection of the spatial index and the semantic index, and the combination result between the geometric level, semantic level and return mode of each instance in the candidate instance set is obtained; The query results corresponding to the query request are generated by combining the results that satisfy the geometric evidence constraints.

8. The data query method for a three-dimensional Gaussian splash scene according to claim 1, characterized in that, The method further includes: The system obtains the data space occupied by the query results in real time and uses the data space occupied to generate the return batch corresponding to the query request. Obtain the query signature of the query request, and generate a token verification result of the query signature using the query identifier and snapshot identifier corresponding to the returned batch; The query result is updated based on the token verification result.

9. A data query system for three-dimensional Gaussian splashing scenarios, characterized in that, The system includes: Atomic Gaussian Cluster Generation Module: Used to construct a spatial hierarchy tree using scene location data in the target 3D Gaussian splash scene, determine a unique master-belonging leaf node for each Gaussian record data in the target 3D Gaussian splash scene based on the spatial hierarchy tree, and use the master-belonging leaf node to perform clustering and merging processing on all Gaussian record data to generate atomic Gaussian clusters; Constraint generation module: used to construct geometric level loads and semantic level data based on the semantic instance graph generated by the atomic Gaussian cluster, integrate the intersection relationship, semantic content, instance graph and level mapping corresponding to the geometric level load and the semantic level data into a scene snapshot, and construct the geometric evidence constraints corresponding to each semantic level; The query result acquisition module is used to parse the query request and generate a query plan that is bound to the scene snapshot. Under the geometric evidence constraint, the module obtains the geometric and semantic query results from the scene snapshot according to the query plan.

10. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing computer-executable instructions that can be executed by the processor, the processor executing the computer-executable instructions to implement the steps of the data query method for a three-dimensional Gaussian splash scene according to any one of claims 1 to 8.

Images