A three-dimensional Gaussian target segmentation method based on bilateral rendering contribution attribution and boundary ambiguity refinement

Abstract

The application provides a three-dimensional Gaussian target segmentation method based on bilateral rendering contribution attribution and boundary ambiguity refinement, and belongs to the technical field of computer vision and computer graphics. The method determines the pixel-level effective rendering contribution of each Gaussian primitive based on the transparency synthesis rule, accumulates the effective contributions of the Gaussian to the target region and the background region respectively to obtain bilateral contribution values, and calculates a boundary ambiguity score according to the bilateral contribution values; the large-support candidate Gaussian is screened in combination with the boundary ambiguity score, the total effective contribution and the spatial scale, the splitting of the spatial support reduction is performed on the large-support candidate Gaussian, the target attribution value of the sub-Gaussian is locally re-estimated, and finally the target Gaussian set is extracted. The application can solve the problem of rough segmentation boundary caused by the mixed support of the cross-boundary Gaussian, control the representation overhead, improve the segmentation precision, and is suitable for three-dimensional scene editing, target removal and the like.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision, computer graphics, 3D scene representation, 3D target segmentation and differentiable rendering technology, and specifically relates to a 3D Gaussian target segmentation method based on bilateral rendering contribution attribution and boundary ambiguity refinement. Background Technology

[0002] 3D Gaussian Splatting (3DGS) represents a 3D scene using multiple explicit anisotropic Gaussian primitives. During rendering, these Gaussian primitives are projected onto the camera's imaging plane to form 2D ellipses, and transparency is composited according to depth order. However, the training objective of standard 3DGS is primarily to interpret the color, transparency, and visibility relationships of multi-view images, rather than to strictly align Gaussian primitive boundaries with semantic object boundaries. Therefore, near object contours, occlusion boundaries, or slender components, the projection support range of a single Gaussian primitive often simultaneously covers both the object region and the background region.

[0003] Existing 3D Gaussian object segmentation techniques based on 2D object supervision typically first acquire multi-view object masks and then elevate 2D object cues to 3D Gaussian units. For example, Chinese invention patent application CN122435137A discloses a single-object 3D model reconstruction method that sets object assignment probabilities for each 3D Gaussian unit, forms a rendering probability map through transparency-weighted accumulation, and optimizes these probabilities using multi-view supervision. However, such schemes usually treat each Gaussian unit as an indivisible single assignment unit, assigning 2D supervision solely based on Gaussian center projection or discrete voting, without directly utilizing the actual effective contributions of Gaussian projection range, transparency, and occlusion order in pixel formation. When a large supporting Gaussian simultaneously forms pixels on both the target and background sides, the optimization of a single assignment value can only be a compromise between foreground and background supervision, failing to change its cross-boundary support at the representation level.

[0004] To improve boundary quality, some existing technologies attempt to decompose or split Gaussians near the boundaries. For example, Chinese invention patent application CN121074068A discloses a 3DGS segmentation method based on boundary adaptive Gaussian splitting, which identifies boundary Gaussians based on gradient consistency, filters small-scale Gaussians, and splits the remaining boundary Gaussians into two scaled sub-Gaussians. Furthermore, existing technologies also include driving Gaussian splitting or denserening based on 2D boundary location, semantic gradient, or general object attribution uncertainty. However, the candidate triggering signals for such schemes are often disconnected from the actual rendering pixel formation process, making it difficult to accurately distinguish between high-impact Gaussians that truly cross semantic boundaries and unstable Gaussians caused by low visibility or mask noise. Simultaneously, directly using reconstruction-driven, scene-wide indiscriminate denserening or splitting can easily lead to excessive growth in the number of Gaussians and abrupt changes in local occlusion relationships.

[0005] Furthermore, even after boundary refinement, difficult Gaussians may still exist due to 2D mask noise, cross-viewpoint inconsistencies, or structural aliasing. Continuing to perform structural decomposition on all difficult Gaussians may cause the number of Gaussians to continue to expand and local representations to become unstable; if a dense classification network is used for post-processing across the entire scene, it will significantly increase computational overhead and may change the already stable Gaussian assignments. Existing technologies lack a mechanism that can perform controlled correction only on a small number of residual difficult candidates while keeping the ordinary stable primitives unchanged. Summary of the Invention

[0006] To address the shortcomings and deficiencies of existing technologies, this invention provides a 3D Gaussian target segmentation method based on bilateral rendering contribution attribution and boundary ambiguity refinement. This method aims to solve the problems in existing 3D Gaussian target segmentation where a single target attribution variable cannot eliminate cross-boundary Gaussian representation support mixing, and where boundary candidate trigger signals are disconnected from actual rendering conflicts. This invention obtains a 3D Gaussian scene and multi-view target masks, and determines the effective rendering contribution of each Gaussian unit to each pixel based on transparency synthesis rendering rules. The key is to accumulate the effective rendering contribution of each Gaussian unit to pixels in both the target and background regions according to the target mask, obtaining the target-side contribution and background-side contribution. A boundary ambiguity score is then calculated based on the relative closeness of these two contributions, thereby directly identifying bilateral contribution conflicts at the representation level. Furthermore, the boundary ambiguity score, total effective contribution, and spatial scale are combined to screen for large support candidate Gaussians, which are then subjected to spatial support reduction splitting with symmetric offset along the principal axis of the maximum scale, scale contraction, and approximate conservation of combined opacity. Subsequently, the target attribution value is locally re-estimated within the rendering influence region of the sub-Gaussian, and the geometric, appearance, and opacity parameters of the remaining Gaussian units are frozen. For small-scale difficult candidate Gaussians that were not split and residual difficult Gaussians that are still unstable after refinement, a candidate correction set is constructed by merging them, and residual correction with amplitude controlled by boundary ambiguity scores is performed. Finally, a set of target Gaussians that is consistent across perspectives, renderable and editable is extracted.

[0007] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0008] This invention provides a 3D Gaussian object segmentation method based on dual-sided rendering contribution attribution and boundary ambiguity refinement, as detailed below:

[0009] A 3D Gaussian scene containing multiple Gaussian primitives and a multi-view target mask are obtained, and a target attribution value is configured for each Gaussian primitive. Based on the transparency compositing rendering rules, the effective rendering contribution of each Gaussian primitive to pixels in each viewpoint is determined. According to the target mask, the effective rendering contribution of each Gaussian primitive to pixels in the target region and pixels in the background region is accumulated respectively to obtain the target-side contribution and background-side contribution of the corresponding Gaussian primitive. The sum of the target-side contribution and the background-side contribution is the total effective contribution. According to the relative closeness between the target-side contribution and the background-side contribution of each Gaussian primitive, the boundary ambiguity score of the Gaussian primitive is calculated. The boundary ambiguity score increases as the closeness between the target-side contribution and the background-side contribution increases. The boundary ambiguity score, the total effective contribution, and the spatial scale of the Gaussian primitives are combined for screening. The selected large-support candidate Gaussians are split by spatial support reduction to generate at least two sub-Gaussians and initialize their parameters. Only the target attribution value of the sub-Gaussians is redefined. During the optimization process, the spatial center, scale, rotation, color parameters, opacity, and target attribution value of all other Gaussian primitives remain unchanged. The target Gaussian set is extracted based on the target attribution value of each Gaussian primitive.

[0010] Furthermore, the calculation rule for the effective rendering contribution is as follows: for any pixel from any viewpoint, the effective rendering contribution of a Gaussian element to that pixel is equal to the local transparency contribution of that Gaussian element at that pixel, multiplied by the cumulative transmittance caused by all Gaussians located in front of that Gaussian element in the transparency composition depth order. This calculation method aligns with the actual rendering logic of 3D Gaussian splashing, and can simultaneously reflect the projection support range of the Gaussian element, its own transparency, and the visibility attenuation caused by in-front occlusion, allowing subsequent supervised attribution to be based on the actual pixel formation weights.

[0011] Furthermore, after determining the effective rendering contribution, an initial optimization step is included: fixing the geometric and appearance parameters of all Gaussian elements, rendering a predicted target mask based on the target attribution value of each Gaussian element and the effective rendering contribution, and optimizing the target attribution value of each Gaussian element according to the difference between the predicted target mask and the multi-view target mask; the step of accumulating the effective rendering contribution of each Gaussian element to the target region pixels and the background region pixels is performed after the initial optimization step. This step, without changing the original scene geometry and appearance, initially fits the target attribution tendency of each Gaussian element, providing an attribution field with basic discriminative power for subsequent two-sided contribution statistics and ambiguity analysis.

[0012] Furthermore, the calculation of the boundary ambiguity score satisfies the following relationship: Take twice the smaller value between the target-side contribution and the background-side contribution, divide by the sum of the target-side contribution, the background-side contribution, and the preset zero-prevention constant, to obtain a boundary ambiguity score ranging from 0 to 1. The preset zero-prevention constant is a very small positive number used to ensure the stability of the numerical calculation and avoid division-by-zero anomalies when the total contribution approaches zero. This score is obtained based on the comparison of the two-sided contributions in actual rendering and can characterize the degree of conflict in the formation of pixels on both the target and background sides during the rendering process.

[0013] Furthermore, the screening process specifically includes: first, selecting highly ambiguous candidate Gaussians whose boundary ambiguity scores are higher than a preset ambiguity threshold and whose total effective contribution is higher than a preset contribution threshold; then, from among the highly ambiguous candidate Gaussians, those with a maximum spatial scale higher than a preset scale threshold are designated as large-support candidate Gaussians, and the rest are designated as small-scale difficult candidate Gaussians. The spatial scale is taken as the maximum value in the anisotropic scale vector of the Gaussian unit, reflecting the spatial extension range of the Gaussian. Through multi-condition joint screening, low-visibility Gaussians with extremely low total contribution and negligible impact on the segmentation results can be eliminated first, and then candidates suitable for different processing strategies can be distinguished based on spatial scale, reducing invalid splitting operations.

[0014] Furthermore, the process of generating child Gaussians and initializing their parameters specifically includes: symmetrically offsetting along the principal axis of the parent Gaussian's maximum scale to generate the spatial centers of two child Gaussians, and proportionally shrinking the anisotropic scale of the child Gaussians; the child Gaussians inherit the rotation parameters, color parameters, and initial target assignment values ​​of the parent Gaussian; the opacities of the two child Gaussians satisfy the approximate conservation rule of combined opacity, making the combined opacity of the two child Gaussians approximately equal to the opacity of the parent Gaussian, thus limiting changes in local occlusion effects and rendering appearance before and after splitting. Here, the approximate conservation of combined opacity means that under the approximate condition of local spatial overlap of the child Gaussians, the equivalent occlusion effect after their superposition is close to the original opacity of the parent Gaussian, not a strict mathematical equation. This design can reduce the disturbance of structural splitting to the local appearance and occlusion relationships of the scene. Splitting along the principal axis of the maximum scale is a conservative approach that does not require additional estimation of the boundary normal, which can significantly reduce the probability of a single child Gaussian simultaneously crossing both sides of the boundary.

[0015] Furthermore, during the segmentation process, no target or background category is pre-assigned to the sub-Gaussian; the target classification of the sub-Gaussian is determined through subsequent re-estimation steps. This design does not require pre-assuming the precise location of the semantic boundary; the final classification of the sub-Gaussian is objectively determined in subsequent steps by multi-view target supervision, reducing bias caused by human prior knowledge.

[0016] Furthermore, the process of redetermining the target attribution value of the sub-Gaussian includes: determining the split rendering influence region based on the effective rendering contribution of the sub-Gaussian; optimizing the target attribution value of the sub-Gaussian only within the rendering influence region; and freezing the geometric parameters, appearance parameters, and opacity of all other Gaussian pixels during the optimization process. Here, the rendering influence region refers to the set of pixels where the sub-Gaussian has a non-zero effective contribution to the pixel. By constraining the optimization range through the influence region, the computational load can be reduced, while avoiding unnecessary perturbations to the unsplit stable Gaussian pixels, thus helping to maintain the stability of the overall scene appearance.

[0017] Furthermore, after redetermining the target assignment value of the sub-Gaussians, a residual correction step is included: the target-side contribution, background-side contribution, and boundary ambiguity score of each Gaussian after splitting are recalculated; residual difficult Gaussians that still meet the preset residual ambiguity and residual contribution conditions after refinement and whose target assignment values ​​are close to the discrimination boundary are selected; the small-scale difficult candidate Gaussians and the residual difficult Gaussians are merged into a candidate correction set; target assignment correction is performed only on Gaussians within the candidate correction set, with the correction magnitude increasing as the boundary ambiguity score of the Gaussian increases; Gaussians outside the candidate correction set retain their original target assignment values. Here, a target assignment value close to the discrimination boundary means that the target assignment value is in the range near the target and background classification thresholds, indicating that the Gaussian's assignment tendency is not yet clear. The correction magnitude is dynamically adjusted according to the degree of ambiguity, allowing Gaussians with stronger boundary conflicts to obtain a larger correction space, while Gaussians with weaker conflicts remain relatively stable, achieving a bounded and controlled correction effect.

[0018] Compared with the prior art, the present invention and its preferred embodiments have at least the following beneficial effects:

[0019] 1. The attribution of target supervision is more accurate. The two-dimensional supervision signal is assigned as weight based on the effective contribution in the actual rendering process. It takes into account the projection range, transparency and occlusion relationship of Gaussian, and can reduce the attribution bias caused by central projection or discrete voting methods.

[0020] 2. The identification of boundary conflicts is more accurate. Based on the actual rendering of the two-sided contribution definition of boundary ambiguity, it can distinguish between high-influence Gaussians that truly participate in the formation of target and background pixels at the same time, and candidates that are unstable due to low visibility and mask noise. This avoids performing invalid splitting on low-contribution Gaussians.

[0021] 3. The boundary refinement process is more controllable. By selecting and adapting candidate Gaussians of different sizes through spatial scale, the large supporting Gaussians are conservatively split to reduce the spatial support. Cross-boundary mixing is eliminated from the representation level. This can improve the accuracy of the segmentation boundary while controlling the growth rate of the number of Gaussians and reducing the disturbance of the splitting to the overall appearance and occlusion relationship of the scene.

[0022] 4. The residual correction mechanism is efficient and lightweight, performing controlled correction only on a small number of difficult candidates. It does not require the deployment of a dense classification network across the entire scene, and the correction magnitude is dynamically adjusted according to the degree of boundary ambiguity. It can compensate for residual boundary errors without changing the stable Gaussian attribution, resulting in lower overall computational overhead.

[0023] 5. The overall system adopts a layered and progressive processing architecture with processing strategies adapted to different types of candidate Gaussian matching. The solution is more robust and can adapt to 3D scenes and target shapes of varying complexity. Attached Figure Description

[0024] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0025] Figure 1 This is a schematic diagram comparing the problem motivation and segmentation effect of an embodiment of the present invention;

[0026] Figure 2 This is a schematic diagram of the overall technical framework of an embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of the method execution flow according to an embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram illustrating the boundary ambiguity-guided Gaussian element refinement principle of an embodiment of the present invention.

[0029] Figure 5 This is a schematic diagram of the candidate-level residual expert correction structure according to an embodiment of the present invention;

[0030] Figure 6 This is a qualitative comparison diagram of the segmentation effects of different methods in embodiments of the present invention;

[0031] Figure 7 This is a schematic diagram of the module-based progressive ablation experiment results according to an embodiment of the present invention. Detailed Implementation

[0032] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0033] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0034] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0035] Combination Figure 1 As can be seen intuitively, when using the conventional 3D Gaussian target segmentation scheme with direct Gaussian selection, the projection support range of the cross-boundary Gaussian unit simultaneously covers both the target area and the background area. A single belonging unit cannot resolve the conflict between the rendering contributions of both sides, and the final extracted target Gaussian set is prone to problems such as background leakage, boundary erosion, and missing local edges of the target.

[0036] Existing solutions still have several limitations in practical applications. In the supervised attribution stage, most solutions only allocate 2D target supervision signals based on Gaussian central projection or cross-view discrete voting, failing to fully consider the Gaussian projection coverage, local transparency, foreground occlusion order, and its actual effective contribution in the final pixel formation process, resulting in inherent biases in attribution accuracy. At the representation level, existing solutions typically treat individual Gaussian units as indivisible single attribution units. Even if intermediate target attribution results are obtained through optimization, the spatial support of the Gaussian still covers both the target and the background. Including such Gaussians in the target set introduces background components, while excluding them may result in the loss of target edges or fine structures. In the boundary candidate recognition stage, existing solutions often use 2D boundary distance, semantic gradient, or ordinary classification uncertainty as the splitting trigger, making it difficult to distinguish between high-influence Gaussians that truly participate in both foreground and background pixel formation and candidates that are unstable due to low visibility and low contribution. This leads to a disconnect between the trigger signal and the actual rendering conflict. In the structural refinement stage, indiscriminate full-scene splitting and densification, or repeated splitting of residual difficult Gaussians, can easily lead to an excessively rapid increase in the number of Gaussians, changes in local occlusion relationships, and additional computational burden. After refinement, if the target assignment is re-optimized for the complete scene, it can easily cause unnecessary changes to the originally stable Gaussians; if local re-estimation is not performed, the sub-Gaussians obtained from the splitting will still inherit the mixed assignment state of the parent Gaussian, failing to achieve the expected refinement effect. In addition, for residual difficult Gaussians in scenarios with low visibility, similar appearance between the target and the background, and overlapping of fine-scale structures, these candidates are often not suitable for further structural splitting. Existing technologies lack a mechanism to perform limited-amplitude correction on residual candidates by comprehensively considering multiple types of evidence without altering the assignment of stable Gaussians.

[0037] Based on the aforementioned limitations of existing technologies, the technical problem this invention aims to solve is, given that 3D Gaussian primitives already possess basic target attribution values ​​and boundary Gaussian splitting techniques have been publicly disclosed, how to accurately distinguish between high-influence cross-boundary Gaussians that truly participate in the formation of pixels on both the target and background sides, and ordinary difficult Gaussians that are unstable only due to low visibility, mask noise, or attribution values ​​approaching the discrimination threshold, relying on the actual transparency synthesis process; and how to conduct controlled local spatial support refinement for high-influence cross-boundary Gaussians, improving regional consistency and boundary stability of the target Gaussian set while controlling the growth of the number of Gaussians and reducing overall scene appearance disturbances. Furthermore, this invention addresses several subdivision issues, including how to filter out low-visibility candidates using the total effective contribution, how to distinguish between large-scale support conflicting Gaussians suitable for structural refinement and small-scale difficult Gaussians unsuitable for further subdivision by combining Gaussian spatial scales, how to re-estimate the target attribution of sub-Gaussians within the subdivision influence region without re-optimizing the entire scene, and how to perform bounded correction controlled by boundary ambiguity only for a small number of residual candidates.

[0038] The present invention will be further described below with reference to the accompanying drawings and through more specific embodiments.

[0039] (a) Basic symbols, data structures and task definition

[0040] Given a pre-reconstructed 3DGS scene, the Gaussian metaset is denoted as... Each Gaussian unit At least including the center of space Anisotropic scale Rotation Opacity and color parameters Given A set of target masks for each training perspective This invention introduces a continuous target attribution value for each Gaussian element. All Gaussian target attribution values ​​form the objectness field for the current target object.

[0041] Because this invention conservatively splits a small number of boundary Gaussians, the final target Gaussian set From the refined Gaussian set Extracted from, but not strictly limited to, the original set. A subset of.

[0042] Table 1. Meaning of Main Parameters

[0043]

[0044] Table 2. Input, Processing Stages, and Output Relationships

[0045]

[0046] (II) Overall Process

[0047] like Figure 2 As shown, this method takes multi-view images, camera pose, pre-trained 3D Gaussian scene, and target mask as input. First, it optimizes target attribution based on the transparency synthesis process, accumulating the effective contributions of each Gaussian on the target and background sides and establishing a boundary ambiguity description. Then, it filters candidates with large support boundaries based on the degree of conflict between the two sides, the total contribution, and the spatial support scale, and performs conservative splitting and local target attribution reassessment on them. For a small number of candidates that are still unstable or unsuitable for further splitting, candidate-level residual expert correction is used. Finally, it extracts the target Gaussian set based on the basic target attribution of the stable Gaussian and the calibrated target attribution of the difficult candidates.

[0048] like Figure 3 As shown, the execution flow of this method can be divided into three progressively advancing processing stages, fully covering the entire chain from input data to target Gaussian output. The first stage is contribution attribution and boundary ambiguity analysis, which sequentially completes four steps: scene and target supervision acquisition, effective contribution calculation and target attribution optimization, two-sided contribution statistics and boundary ambiguity modeling, and boundary candidate screening. Finally, it outputs the target-side contribution and background contribution statistics of each Gaussian, and filters out three types of intermediate results: highly ambiguous candidate set, large support candidate set to be split, and small-scale difficult candidate set. The second stage is boundary Gaussian refinement and local reassessment. First, the boundary Gaussian conservative split is performed on the candidates to be split, and sub-Gaussians are generated by deterministic symmetric offset and scale contraction, and the parameters are initialized with the opacity approximate conservation rule. Then, local target attribution reassessment is performed within the split influence range. During the update process, only the target attribution value of the sub-Gaussians is adjusted, while the other geometric, appearance, and opacity parameters remain fixed. The third stage is residual candidate correction. First, small-scale difficult candidates and residual candidates that are still unstable after refinement are merged to form the final correction candidate set. Then, three types of features, namely boundary, appearance and structure, are constructed for each candidate. Bounded residual update is performed through an architecture of multiple micro-experts combined with Top-k routing. Finally, the correction process is completed and the target set is extracted, outputting a 3D target segmentation result with more stable boundary quality.

[0049] like Figure 4As shown, Gaussian elements distributed across the target boundary are simultaneously affected by the monitoring signals from both the target and background regions. When the effective contributions of the Gaussian to both sides are relatively close, the boundary ambiguity score is at a high level. Before performing structural refinement, candidates to be split need to be screened through three conditions. Only Gaussian elements that simultaneously meet the following criteria—boundary ambiguity higher than a preset threshold, total effective contribution higher than a preset threshold, and maximum spatial scale higher than a preset threshold—will be selected for splitting. The parent Gaussian that meets the conditions generates child Gaussians using conservative splitting rules. By proportionally shrinking the spatial scale and setting an approximately conserved opacity parameter, the split child Gaussians have a smaller spatial support range while maintaining a local occlusion effect that is basically consistent with that before splitting. After splitting, the support range of most child Gaussians is biased towards the target side or the background side, and the classification clarity is improved. A small number of small-scale residual candidates that still have classification ambiguity are transferred to the subsequent residual expert correction stage for further processing.

[0050] like Figure 5 As shown, residual candidate correction adopts a candidate-level processing mode. The correction input set is composed of small-scale difficult candidates and refined residual difficult candidates. The correction operation only applies to Gaussian elements within this set and does not affect other stable Gaussians. The correction architecture includes three parallel lightweight micro-expert branches, corresponding to boundary, appearance, and structural feature inputs, respectively. Each branch independently outputs the correction prediction score for the corresponding dimension through a multilayer perceptron. The outputs from the three micro-experts are processed by a Top-k routing module, which selects the experts with the highest weights for weighted fusion to obtain the comprehensive correction result. Finally, a bounded residual update method is used to adjust the target assignment value of the Gaussian. The correction magnitude is dynamically controlled by the degree of boundary ambiguity of the Gaussian elements; the higher the ambiguity, the larger the allowed correction space. After calibration, target candidates and background candidates are distinguished according to the assignment threshold.

[0051] (III) S101: Acquiring a 3D Gaussian scene and multi-view target supervision

[0052] The system acquires a pre-trained 3D Gaussian scene, multi-view training images, camera intrinsic and extrinsic parameters or camera poses corresponding to each training image, and target masks for each training viewpoint. The target mask specifies the object to be extracted and can be generated by the 2D object segmentation model based on points, bounding boxes, masks, or semantic cues.

[0053] In this embodiment of the invention, the scene contains 40 to 120 multi-view images, with the training viewpoint and the evaluation viewpoint divided in an 8:2 ratio. The training viewpoint mask is used for target attribution optimization, contribution statistics, candidate selection, local refinement, and residual correction; the evaluation viewpoint mask is only used to calculate evaluation metrics and does not participate in the above optimization processes.

[0054] (iv) S102: Optimize the target attribution field based on the effective contribution of transparency synthesis

[0055] For perspective The pixels below ,set up This is the set of Gaussian elements involved in rendering this pixel. For pixels The effective contribution is determined by its local alpha contribution at that pixel and the cumulative transmittance before the light reaches that Gaussian:

[0056]

[0057] in, Gauss From the perspective pixels The contribution of local transparency generated at that location, Indicates the position in the transparency composition order The cumulative transmission attenuation caused by the Gaussian in front express lie in Ahead. Effective contribution. It also reflects the projection support, transparency, occlusion order, and actual visibility.

[0058] Based on the target attribution value of each Gaussian Render the 3D target attribution field onto a 2D mask plane:

[0059]

[0060] In the target attribution stage, the spatial center, scale, rotation, color, and opacity of the original 3DGS are kept unchanged. Only the target attribution values ​​of each Gaussian are optimized, and the multi-view target mask reconstruction loss is minimized.

[0061]

[0062] Using unnormalized transparency to synthesize contributions can reduce the impact of low opacity or occluded Gaussians on target supervision, thus avoiding the uniform distribution of target supervision for the same pixel to all visible Gaussians.

[0063] Preferably, in this embodiment of the invention, the objectness is initialized to 0.5, and Adam is used for 600 iterations of optimization with a learning rate of [missing information]. .

[0064] (v) S103: Accumulate the contributions from the target side and the background side and determine the boundary ambiguity.

[0065] A single target assignment value only indicates the final assignment tendency and cannot directly indicate whether the Gaussian scalar actually forms pixels on both the target and background sides simultaneously. To explicitly characterize bilateral supervised conflict, the effective contribution of each Gaussian scalar to the target and background regions is accumulated separately based on the target mask:

[0066]

[0067] in, Gauss The cumulative effective contribution to the formation of pixels in the target region. This indicates its cumulative effective contribution to the formation of background pixels. This indicates the overall effectiveness of Gaussian participation in the current objective monitoring.

[0068] The boundary ambiguity score is defined based on the proximity of the two-sided contributions:

[0069]

[0070] when and When all values ​​are relatively large and close, A value close to 1 indicates that the Gaussian gas significantly contributes to the formation of pixels on both the target and background sides through the actual rendering path; when the contribution is significantly biased to one side, A value close to 0 indicates that the target assignment of the Gaussian is relatively stable.

[0071] This boundary ambiguity differs from that caused by The former represents the general classification uncertainty. It directly stems from the two-sided contribution of the target mask to the Gaussian along the transparency synthesis path, while the latter only indicates that the target's classification value is close to the midpoint of the discrimination.

[0072] (vi) S104: Joint ambiguity, contribution and spatial support for screening candidate sets

[0073] High boundary ambiguity does not necessarily indicate that the Gaussian will significantly affect the target boundary. For example, a Gaussian with low opacity or strong occlusion may only produce a small two-sided contribution. To exclude such low-impact candidates, this method combines boundary ambiguity, total effective contribution, and Gaussian spatial scale for candidate selection:

[0074]

[0075] in, A highly ambiguous but fully contributing set of candidates; for Gaussian structures with large spatial support are suitable for reducing cross-boundary coverage through structural refinement. for Difficult candidates that are too small at the mesoscale or unsuitable for further splitting will be left for subsequent residual correction.

[0076] Preferably, in this embodiment of the invention, the candidate screening threshold is adopted. , and The cumulative contribution and the maximum Gaussian scale are normalized within the current scenario.

[0077] (vii) S105: Perform conservative splitting on highly ambiguous, large-support Gaussian.

[0078] For each The parent Gaussian is replaced with two smaller child Gaussians supported by local spatial support. This process does not pre-designate the two child Gaussians as the target side or the background side, nor does it claim to recover the precise 3D semantic boundary. Instead, it first decomposes a large cross-boundary support interpretation unit into more local representation units, and then redetermines its target attribution by multi-view target supervision.

[0079] Let the covariance matrix of the father Gauss be... To ensure refinement stability and reproducibility, a deterministic symmetric migration is performed along the principal axis of the parent Gaussian maximum scale, and the scale is shrunk:

[0080]

[0081] in, The axis corresponding to the maximum scale of the parent Gaussian. This is the direction of the principal axis in the world coordinate system. Control the offset of the two child Gaussians relative to the parent center. is the scaling factor. The maximum scale principal axis serves only as a conservative approximation that does not require additional boundary normal estimation; its role is to reduce the probability that a single sub-Gaussian simultaneously crosses both sides of the boundary.

[0082] To reduce abrupt changes in appearance and occlusion after splitting, the two child Gaussians inherit the rotation, color, and initial target assignment from the parent Gaussian, and employ an initialization that approximates combined opacity conservation:

[0083]

[0084] Under the local overlap approximation, the combined opacity of the two sub-Gaussians is close to the opacity of the parent Gaussian, so that structural refinement mainly changes the local spatial support at the boundary, rather than uncontrollably increasing local occlusion. Preferably, the embodiments of the present invention employ... and .

[0085] (viii) S106: Reassess the local target attribution within the affected area of ​​the split.

[0086] The updated set is obtained after splitting. ,in This creates a new set of sub-Gaussians. During the local revaluation phase, the geometry, color, and opacity of all Gaussians remain fixed, and the target assignment of the newly generated sub-Gaussians is updated only.

[0087] The splitting influence weight is determined based on the effective contribution of the newly generated sub-Gaussian to each pixel, and the local mask reconstruction loss is calculated within this influence region. At the same time, a regularization term is used to limit the deviation of the sub-Gaussian target assignment from its inherited initial value.

[0088]

[0089] In the above formula, the first term ensures that the predicted target mask within the split influence region is consistent with the multi-view target mask; the second term reduces over-updating caused by 2D mask noise, insufficient local visibility, or occlusion. Since geometry, color, and opacity remain frozen, local re-evaluation does not re-optimize the entire 3DGS scene.

[0090] In actual calculations, the set of pixels whose influence weight is greater than zero (or greater than a preset minimum threshold) can be identified as the rendering influence area.

[0091] Preferably, in this embodiment of the invention, the new sub-Gaussian target assignment is optimized for 300 iterations, with a learning rate of [missing information]. Regular weights .

[0092] (ix) S107: Forming a residual candidate set

[0093] After the local target attribution reassessment, in The target-side contribution, background-side contribution, total contribution, and boundary ambiguity are recalculated. For Gaussians that still have high ambiguity, sufficient contribution, and target attribution close to the discrimination boundary, geometric splitting is no longer performed; instead, residual candidates are formed.

[0094] in, This represents a residual Gaussian that, even after refinement, is still difficult to stably classify into the target or background. From the original small-scale difficult candidate and refined residual candidates Together they constitute. Preferably, the embodiments of the present invention employ... .

[0095] This set construction prioritizes resolving large support cross-boundary conflicts through representation structure refinement, while difficult Gaussians with small scale, low visibility, similar appearance, or affected by 2D mask noise are entered into the candidate-level correction, avoiding representation fragmentation caused by repeated splitting.

[0096] (x) S108: Construct candidate features and perform lightweight residual expert correction

[0097] The refined basic objectives are categorized as follows: .like If empty, skip candidate corrections; if Non-empty, constructing only three sets of low-dimensional features—boundary, appearance, and structure—for candidate Gaussians:

[0098]

[0099] Boundary features include boundary ambiguity, normalized total contribution, normalized distance to the target mask boundary, and uncertainty in the determination of the basic target attribution; appearance features include local color differences and RGB reconstruction residuals; structural features include normalized scale, opacity, visibility ratio, and local neighborhood density. In one implementation, the above candidate features are calculated in the following manner.

[0100] 1. A specific calculation method for candidate features

[0101] (1) Normalized distance to the target mask boundary

[0102] For the target mask of training viewpoint v Extract the set of boundary pixels of the target mask. The Euclidean distance from pixel p to the target mask boundary, calculated according to the preset boundary bandwidth. The normalized pixel-level distance and the Gaussian-level boundary distance feature obtained by aggregating effective rendering contributions are calculated according to the following formulas:

[0103]

[0104] in, Gausky element The effective rendering contribution of pixel p in viewpoint v is ε, which is a positive number to prevent the denominator from being zero. This calculation considers not only the Gaussian center projection position, but also the Gaussian projection range, transparency, occlusion order, and actual visible contribution.

[0105] (2) Local color differences

[0106] Based on the effective rendering contribution of Gaussian elements to pixels in the target and background regions, the weighted RGB color mean values ​​for the target and background sides are calculated, and the average L1 distance between the two color means is taken as the local color difference.

[0107]

[0108] in, This represents the RGB color of pixel p in the training viewpoint v. and Representing Gaussian elements respectively The cumulative effective contribution to the formation of pixels in the target and background regions; dividing by 3 indicates averaging the absolute differences of the three RGB color channels.

[0109] (3) RGB reconstruction residual

[0110] Let the real RGB image of the training viewpoint v be denoted as The rendered image of the complete 3D Gaussian scene from this viewpoint is denoted as First, the pixel-level RGB average L1 reconstruction error is calculated. Then, based on the effective rendering contribution of Gaussian primitives, the pixel-level residuals are aggregated to the Gaussian level.

[0111]

[0112] The RGB reconstruction residual is used to characterize the reliability of the appearance reconstruction of the region formed by the candidate Gaussian; the larger the residual, the more unstable the appearance or geometric reconstruction of the corresponding region may be.

[0113] (4) Visible proportions

[0114] First, calculate the Gaussian element. Cumulative effective contribution under training perspective v .set up For Gaussian Yuan The set of training viewpoints whose projection intersects with the image plane is used to determine whether each viewpoint is effectively visible in each training viewpoint using a threshold relative to the maximum single-viewpoint contribution of the Gaussian viewpoint, and the visibility ratio is calculated:

[0115]

[0116] Where κ∈(0,1) is the preset visibility judgment coefficient. This is used for binary visibility labeling. The denominator uses the number of training viewpoints whose actual Gaussian projection enters the image plane, rather than the total number of training viewpoints, to avoid incorrectly counting viewpoints where the Gaussian projection does not enter the camera's view frustum as invisible.

[0117] (5) Local neighborhood density

[0118] In one embodiment of the present invention, the K-nearest neighbor inverse bulk density is employed. Let... For Gorsky Yuan Space Center The Euclidean distance to the Kth nearest neighbor Gaussian center, where K is a preset positive integer, is used to calculate the local neighborhood density according to the following formula:

[0119]

[0120] The K-nearest neighbor form can adapt to changes in Gaussian distribution density in different regions of the same scene; the denser the local region, the smaller the distance of the K-nearest neighbor, and the greater the corresponding neighborhood density.

[0121] (6) Scene-based normalization

[0122] For scalar features that require in-scene normalization In one embodiment of the invention, minimum-maximum normalization is employed:

[0123]

[0124] in, and These are the minimum and maximum values ​​of the corresponding features in the current scene or the current candidate set, respectively, and the normalization result can be further truncated to [0,1]. In other implementations, mean-standard deviation normalization or normalization based on preset quantiles can also be used, as long as the corresponding features between different candidate Gaussians have comparable numerical scales.

[0125] The three sets of features are input into three lightweight micro-experts: boundary, appearance, and structure. The routing module generates routing probabilities based on the candidate complete features, selects the Top-k experts, and performs a weighted fusion of the selected expert outputs.

[0126]

[0127] Preferably, in this embodiment of the invention, each micro-expert employs a two-layer MLP with a hidden dimension of 32. Routing uses The micro-expert output is not a direct replacement for the basic target attribution, but rather generates candidate correction scores.

[0128] Determine the correction gating based on boundary ambiguity and perform bounded residual updates:

[0129]

[0130] in, The correction magnitude increases with the residual boundary ambiguity. Controlling the overall residual intensity. Gaussians outside the candidate set retain their basic target assignments, avoiding the impact of the lightweight correction module on ordinary stable Gaussians. Preferably, this embodiment of the invention employs... .

[0131] (xi) S109: Train candidate correction modules and extract the final target Gaussian set

[0132] Candidate correction training phase freeze The geometry, color, opacity, and fundamental target attribution are used to optimize only the expert and routing parameters. To restrict supervision to the candidate's actual influence region, candidate influence weights are constructed based on the effective contribution of the candidate Gaussian, and the following training objective is used:

[0133]

[0134] The first loss term constrains the calibration prediction mask within the candidate influence area to be consistent with the multi-view target mask; the second term restricts the calibration target affiliation from deviating too far from the basic target affiliation; the third term is a weak route balancing regularization to avoid all candidates degenerating to the same expert; and the fourth term is an expert parameter regularization. This represents the set of parameters for the micro-expert and routing module to be optimized.

[0135] Preferably, in this embodiment of the invention, the candidate correction is optimized through 400 iterations, with a learning rate of... ,use , and .

[0136] During the final extraction, stable Gaussians outside the candidate set are determined using the base target assignment value, while difficult Gaussians within the candidate set are determined using the calibrated target assignment value.

[0137]

[0138] As a preferred embodiment of the present invention, the following is adopted: and .final It can render target objects individually, or remove them from the entire scene to support object removal, or serve as a unit of operation for object-level editing, copying, transforming, reassembling, and 3D completion.

[0139] In summary, the method provided in this embodiment alleviates the segmentation accuracy problem caused by cross-boundary Gaussian support mixture from the representation level by steps such as contribution attribution, conflict modeling, local representation refinement, residual candidate correction, and target set extraction.

[0140] This method assigns two-dimensional supervision signals based on the actual effective rendering contribution during the transparency synthesis process, fully considering the Gaussian projection range, opacity, occlusion order, and visibility. Compared to conventional methods such as center projection or discrete voting, the accuracy of supervised attribution is higher. By accumulating the effective contributions of Gaussians to both sides of the target and background and calculating the boundary ambiguity score, the boundary problem is transformed into a two-sided contribution conflict at the Gaussian representation level. This method can accurately distinguish between high-influence Gaussians that truly cross semantic boundaries and ordinary candidates with uncertain attribution, making boundary identification more targeted.

[0141] Building upon this foundation, this scheme employs a three-dimensional approach—boundary ambiguity, total effective contribution, and spatial scale—to screen candidates. This not only eliminates invalid candidates with low visibility and low contribution but also distinguishes between large-support Gaussians suitable for structural refinement and small-scale difficult Gaussians unsuitable for splitting. Only a small number of qualified candidates are split, effectively controlling the growth of the number of Gaussians. The splitting adopts a deterministic conservative design, reducing the spatial support of sub-Gaussians through center offset and scale contraction. Combined with parameter inheritance and an initialization rule that approximates the conservation of opacity, the refinement process primarily adjusts local spatial support, minimizing disturbances to the overall scene appearance and occlusion relationships. Subsequent target attribution re-evaluation is strictly limited to the rendering influence area of ​​the sub-Gaussians, freezing the geometry, appearance, and opacity parameters of the remaining Gaussians throughout the process. Simultaneously, offset constraints are introduced to suppress mask noise interference, avoiding appearance drift and computational redundancy caused by full-scene re-optimization.

[0142] For residual candidates that remain unstable after refinement, this solution merges them with the initially selected small-scale difficult candidates, constructing multi-dimensional features and performing corrections only for a small number of candidates, eliminating the need to deploy a dense classification network across the entire scene. The correction process integrates multi-class feature evidence through multi-micro-expert and sparse routing, and the correction magnitude is dynamically controlled by the degree of boundary ambiguity, avoiding direct replacement of the base attribution value. This allows for compensation of residual errors within a limited adjustment range while ensuring the stable Gaussian attribution state remains undisturbed. The final extracted target Gaussian set can independently perform rendering, deletion, and editing operations, directly adapting to various application scenarios such as target removal, scene editing, and object reorganization.

[0143] To verify the effectiveness of this method, comparative experiments and ablation analyses were conducted on real 3D scene datasets. All experiments were carried out based on a unified hardware and software environment and evaluation protocol to ensure that the comparison results were fair and reproducible.

[0144] The experiment was conducted using the Ubuntu operating system, Python, and PyTorch framework. A dedicated graphics processor was configured on the server to accelerate computation, and all input images were uniformly adjusted to a fixed resolution. All 3D Gaussian scenes used in the experiment underwent standard pre-training, achieving stable geometry, color, opacity, and camera parameters after a sufficient number of iterations. Multi-view images were divided into training and evaluation views at a fixed ratio. The target mask corresponding to the evaluation view was not involved in any optimization process and was only used for final result verification.

[0145] After generating a target mask corresponding to the specified target viewpoint, all Gaussian target attribution values ​​are initialized to a unified initial value. The original geometric and appearance parameters of the 3D Gaussian remain unchanged, and three stages of processing are executed sequentially. Each stage uses the Adam optimizer, with different update objects and iteration counts: The first stage is contribution attribution and target attribution optimization, optimizing the target attribution values ​​of all Gaussians, with a total of 600 iterations; the second stage is boundary Gaussian refinement and local reestimation, updating only the target attribution values ​​of the sub-Gaussians generated by the splitting, with a total of 300 iterations; the third stage is residual candidate correction, updating only the parameters of the micro-experts and routing modules, with a total of 400 iterations.

[0146] The experiment used two datasets for validation. One dataset was the BAR-Scenes dataset, which contains 10 real-world multi-view scenes, each with dozens to hundreds of captured viewpoints. The other dataset was the publicly available Inpaint360GS dataset, used for external validation and not involved in any parameter selection or scene-by-scene tuning. All comparison methods used the same pre-trained 3D Gaussian mask, target definition, viewpoint segmentation, and reference mask to ensure consistent benchmarks.

[0147] The evaluation metrics include three categories: mean intersection-over-union ratio (MIRR) and mean accuracy, which measure the regional consistency of the target segmentation results and reflect the overall segmentation accuracy; and boundary intersection-over-union ratio, which specifically measures the overlap quality of the target boundary regions and focuses on the boundary effects that this method prioritizes for optimization. All metrics are calculated based on a unified reference mask, objectively reflecting the degree of agreement between the segmentation results of each method and the reference standard.

[0148] In terms of actual segmentation results, this method outperforms existing similar schemes in both region consistency and boundary quality. Figure 6 As shown, in the same scenario, the segmentation results of existing methods are prone to problems such as boundary blurring, background residue, or missing target edges. The target Gaussian set extracted by this method has cleaner boundaries after rendering, and the contours fit the actual shape of the target better. The problems of background leakage and boundary erosion are significantly alleviated.

[0149] Table 3. Verification of Regional Consistency and Boundary Quality

[0150]

[0151] Quantitative comparison results show that, under a unified reference mask protocol, our method achieves optimal performance across all three metrics on both datasets. On the BAR-Scenes dataset, our method improves the boundary intersection-union ratio by 2.0 percentage points compared to the existing best comparison scheme; on the Inpaint360GS dataset, the boundary intersection-union ratio is improved by 1.8 percentage points, indicating that the technical improvement of our method mainly affects the target boundary region, consistent with the design expectations.

[0152] To verify the actual function of each core module, a progressive ablation experiment was conducted, gradually adding different functional modules and recording performance changes. The results are as follows: Figure 7 As shown, without the three core modules of this invention, the boundary intersection-union ratio of the basic scheme is 82.7; after adding the contribution attribution module, the boundary intersection-union ratio increases to 83.8, indicating that based on the effective rendering of the contribution allocation supervision signal, the allocation accuracy of target supervision can be improved and the attribution bias caused by the center projection method can be reduced.

[0153] Table 4 Module Progression Effect Verification

[0154]

[0155] COA improved BIoU from 82.7 to 83.8, indicating that the transparency synthesis contribution attribution improved the target supervision assignment; BAR further improved to 86.0, which is the main source of boundary quality; REC improved BIoU to 86.5, providing compensation for a small number of residual difficult candidates.

[0156] Building upon this foundation, the addition of a boundary ambiguity refinement module further improved the boundary intersection-union ratio (IU) to 86.0, which is the main source of the improved boundary quality. This verifies the core role of conservative splitting and local reassessment in resolving cross-boundary support mixing. Finally, the addition of a residual candidate correction module resulted in an IU of 86.5, indicating that controlled correction for a small number of residual difficult candidates can further compensate for boundary errors and improve the final segmentation results without increasing excessive overhead.

[0157] From the perspectives of computational complexity and representation stability, the processing overhead and scene perturbation of this method are both at a low level. Throughout the entire process, only the contribution statistics stage requires traversing the entire scene; subsequent structural refinement and residual correction only apply to a small number of candidate Gaussians: approximately 1.6% of the total number of Gaussians in the scene are involved in structural refinement, and approximately 2.3% are involved in residual correction. Ultimately, the average increase in the number of Gaussians across the entire scene is only 1.6%, far lower than that of full-scene densification schemes, effectively controlling the representation expansion problem.

[0158] Table 5. Verification of Complexity and Representation Stability

[0159]

[0160] Appearance perturbation tests show that after performing conservative splitting and local revaluation, the peak signal-to-noise ratio of the full-scene rendering only decreased by 0.14dB compared to the original scene. This indicates that the structural refinement process did not significantly damage the overall rendering appearance of the scene, and the side effects of local adjustments were effectively limited, thus maintaining the rendering quality of the original scene.

[0161] Table 6 Parameter Sensitivity

[0162]

[0163] Sensitivity tests on the core screening thresholds show that the segmentation performance remains relatively stable when the ambiguity threshold, contribution threshold, and scale threshold are adjusted within a certain range. Too low a threshold expands the candidate range and increases the number of splits, but fails to continuously improve boundary quality; too high a threshold makes the splitting strategy overly conservative, missing some true cross-boundary Gaussian growth. The default parameter configuration achieves a good balance between boundary quality and Gaussian growth, adapting to most common 3D scenes.

[0164] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0165] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A 3D Gaussian object segmentation method based on dual-sided rendering contribution attribution and boundary ambiguity refinement, characterized in that, Includes the following steps: Obtain a 3D Gaussian scene containing multiple Gaussian primitives and a multi-view target mask, and configure target attribution values ​​for each Gaussian primitive; Based on the transparency compositing rendering rules, determine the effective rendering contribution of each Gaussian primitive to pixels at each viewpoint. Based on the target mask, the effective rendering contribution of each Gaussian element to the target region pixels and the background region pixels is accumulated respectively to obtain the target-side contribution and background-side contribution of the corresponding Gaussian element. The sum of the target-side contribution and the background-side contribution is the total effective contribution; The boundary ambiguity score of each Gaussian cell is calculated based on the relative proximity between the target-side contribution and the background-side contribution; the boundary ambiguity score increases as the proximity between the target-side contribution and the background-side contribution increases. The joint boundary ambiguity score, total effective contribution, and spatial scale of Gaussian elements are used for screening. The selected large-support candidate Gaussians are split by spatial support reduction to generate at least two sub-Gaussians and initialize their parameters. Only the target assignment value of the sub-Gaussian is redefined, while the spatial center, scale, rotation, color parameters and opacity of all other Gaussian elements remain unchanged during the optimization process; The target Gaussian set is extracted based on the target attribution value of each Gaussian element.

2. The 3D Gaussian target segmentation method based on dual-sided rendering contribution attribution and boundary ambiguity refinement according to claim 1, characterized in that: The calculation rule for the effective rendering contribution is as follows: for any pixel under any viewpoint, the effective rendering contribution of the Gaussian element to the pixel is equal to the local transparency contribution of the Gaussian element at the pixel, multiplied by the cumulative transmittance caused by all Gaussians in front of the Gaussian element in the transparency composition depth order.

3. The 3D Gaussian target segmentation method based on dual-sided rendering contribution attribution and boundary ambiguity refinement according to claim 1, characterized in that: Following the determination of valid rendering contributions, an initial optimization step is also included: With all Gaussian elements' geometric and appearance parameters fixed, a predicted target mask is obtained by rendering based on the target attribution value of each Gaussian element and the effective rendering contribution. The target attribution value of each Gaussian element is then optimized based on the difference between the predicted target mask and the multi-view target mask. The step of accumulating the effective rendering contribution of each Gaussian pixel to the target region pixels and the background region pixels is performed after the initial optimization step.

4. The 3D Gaussian target segmentation method based on dual-sided rendering contribution attribution and boundary ambiguity refinement according to claim 1, characterized in that: The calculation of the boundary ambiguity score satisfies the following relationship: take twice the smaller value of the target side contribution and the background side contribution, divide it by the sum of the target side contribution, the background side contribution and the preset zero-prevention constant, and obtain the boundary ambiguity score with a value range of 0 to 1.

5. The 3D Gaussian target segmentation method based on dual-sided rendering contribution attribution and boundary ambiguity refinement according to claim 1, characterized in that: The screening process specifically includes: First, select highly ambiguous candidate Gaussians whose boundary ambiguity scores are higher than a preset ambiguity threshold and whose total effective contribution is higher than a preset contribution threshold; Then, from the highly ambiguous candidate Gaussians, those with the largest spatial scale exceeding the preset scale threshold are selected as large support candidate Gaussians, and the rest are selected as small scale difficult candidate Gaussians.

6. The 3D Gaussian target segmentation method based on dual-sided rendering contribution attribution and boundary ambiguity refinement according to claim 1, characterized in that: The process of generating a sub-Gaussian and initializing its parameters specifically includes: The spatial centers of two child Gaussians are generated by symmetrically offset along the principal axis of the parent Gaussian at its maximum scale, and the anisotropic scale of the child Gaussians is proportionally shrunk. The child Gaussian inherits the rotation parameters, color parameters, and initial target assignment values ​​of the parent Gaussian. The opacities of the two child Gaussians satisfy the approximate conservation rule of combined opacity, making the combined opacity of the two child Gaussians approximately equal to the opacity of the parent Gaussian, thus limiting the changes in local occlusion effects and rendering appearance before and after splitting.

7. The 3D Gaussian target segmentation method based on dual-sided rendering contribution attribution and boundary ambiguity refinement according to claim 1, characterized in that: During the splitting process, no target or background category is pre-assigned to the sub-Gaussian; the target affiliation of the sub-Gaussian is determined through subsequent re-estimation steps.

8. The 3D Gaussian target segmentation method based on dual-sided rendering contribution attribution and boundary ambiguity refinement according to claim 1, characterized in that: The process of redetermining the target attribution value of the sub-Gaussian includes: determining the split rendering influence region based on the effective rendering contribution of the sub-Gaussian, optimizing the target attribution value of the sub-Gaussian only within the rendering influence region; and freezing the geometric parameters, appearance parameters, and opacity of all other Gaussian elements during the optimization process.

9. A 3D Gaussian target segmentation method based on dual-sided rendering contribution attribution and boundary ambiguity refinement according to claim 8, characterized in that: After redetermining the target assignment value of the sub-Gaussian, a residual correction step is also included: After recalculating the target-side contribution, background-side contribution, and boundary ambiguity score of each Gaussian after splitting, the residual difficult Gaussians that still meet the preset residual ambiguity condition and preset residual contribution condition after refinement and whose target assignment value is close to the discrimination boundary are selected. The small-scale difficult candidate Gaussian is merged with the residual difficult Gaussian into a candidate correction set. Target assignment correction is performed only on Gaussian within the candidate correction set. The correction magnitude increases as the boundary ambiguity score of the Gaussian increases. Gaussian outside the candidate correction set retains its original target assignment value.

Images