An artificial intelligence-based three-dimensional design collaborative annotation intelligent aggregation method and system

Abstract

The application discloses a three-dimensional design collaborative annotation intelligent collection method and system based on artificial intelligence, and relates to the field of computer-aided design collaboration. The method comprises the following steps: obtaining an original annotation set containing three-dimensional anchor points and text content generated in a collaborative review stage; adopting a space-semantic coupling clustering algorithm to comprehensively calculate the spatial proximity and text semantic similarity between annotation objects, and merging the annotation objects into a plurality of annotation clusters; filling the annotation cluster content into a preset prompt word template, using a large language model to remove duplicates and generate an abstract, and generating a structured rework task list; and establishing a linkage mapping between the task and the best observation view in a collaborative interface. The application solves the problems of annotation fragmentation, high redundancy and lack of execution force in traditional three-dimensional collaboration, automatically converts scattered review opinions into clear engineering instructions through AI intelligent arrangement, and significantly improves the design iteration efficiency.

Description

Technical Field

[0001] This invention relates to the fields of computer-aided design (CAD), computer-supported collaborative work (CSCW), and artificial intelligence, and particularly to a method and system for intelligent summarization of collaborative annotations in 3D design based on a large language model. Background Technology

[0002] In fields such as industrial design, architectural engineering, and game development, the design review of 3D models is a complex collaborative process involving multiple people. Designers, engineers, project managers, and clients typically need to review the models in a shared 3D environment and provide feedback.

[0003] Existing collaborative review tools (such as AutodeskViewer and Figma3D plugins) allow users to annotate model surfaces and leave comments, creating "3D annotations." However, as the number of reviewers increases, this unstructured feedback mechanism exposes serious efficiency problems:

[0004] First, there's the issue of fragmented and redundant information. For the same design flaw (e.g., "the door handle is too high"), five different reviewers might leave five differently worded comments in the same location. When making revisions, designers need to click, read, and manually filter out duplicate information one by one, making it easy to miss or repeat the same work.

[0005] Second, there is a lack of execution. Annotations are often colloquial and emotional (such as "something feels off here"), lacking clear engineering guidance. Designers often need to spend a lot of time figuring out the intentions and translating them into specific "rework tasks."

[0006] Third, spatial positioning is chaotic. When the model is densely covered with hundreds of annotation points, visual occlusion is severe, making it difficult for managers to quickly identify which problems belong to "structural" and which belong to "material".

[0007] Therefore, there is an urgent need for an intelligent processing method that can automatically organize, deduplicate, and refine these three-dimensional spatial annotations, and transform them into a clear, structured task list. Summary of the Invention

[0008] This invention aims to solve the aforementioned technical problems by proposing an intelligent aggregation method that combines "spatial-semantic coupled clustering" with large language model summarization. This method not only considers the textual content of the annotations but also deeply utilizes the positional relationships of the annotations in three-dimensional space, achieving high-precision feedback aggregation and task generation.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] An AI-based intelligent summary method for collaborative annotation in 3D design includes the following steps:

[0011] (1) Obtain the original annotation set generated by the three-dimensional design model in the collaborative review stage; the original annotation set contains multiple discrete annotation objects, each annotation object contains at least three-dimensional anchor point coordinates, camera view parameters and unstructured text content; the camera view parameters include viewpoint position, line of sight direction and field of view angle;

[0012] The original annotation set is subjected to spatial-semantic coupling clustering to generate several annotation clusters. The clustering process specifically includes: extracting the three-dimensional anchor point coordinates of every two annotation objects, calculating the Euclidean distance, and determining that they are spatially adjacent if the distance is less than a preset spatial threshold; converting the text content of each annotation object into a high-dimensional semantic vector using a word embedding model, calculating the cosine similarity between the vectors, and determining that they are semantically related if the similarity is greater than a preset semantic threshold; performing view frustum overlap detection: determining whether there is an overlapping region between the camera view frustums of two annotation objects, and prohibiting them from being merged into the same annotation cluster even if the spatial distance meets the condition if the overlap region volume ratio is less than a preset threshold; constructing an annotation relationship graph, connecting annotation objects that simultaneously meet the spatial proximity condition, the semantic related condition, and pass the view frustum detection into a connected subgraph, and each connected subgraph constitutes an annotation cluster.

[0013] Construct a prompt word template that includes role settings and output constraints, serialize all text content and corresponding position information in each annotation cluster and fill it into the prompt word template to generate a large language model input vector;

[0014] The input vector is fed into a pre-trained large language model for intent understanding and summary generation to obtain a structured list of rework tasks;

[0015] In the collaborative design interface, the rework task list is associated with the 3D design model and a navigation entry is provided to locate the corresponding annotation cluster space with one click.

[0016] In response to the user selecting a task in the rework task list, the system automatically parses the annotation cluster ID associated with the task; calculates a weighted average optimal viewing angle based on the camera viewpoint parameters of all annotation objects contained in the annotation cluster; and controls the 3D viewport to smoothly transition to the optimal viewing angle.

[0017] (2) Further, the construction of the prompt word template, which includes character settings and output constraints, specifically includes:

[0018] Set the model's role as either a design director or an engineering manager;

[0019] Configure deduplication instructions to instruct the model to identify and merge feedback that contains semantic duplicates;

[0020] Configure tone conversion instructions to instruct the model to convert colloquial comments into imperative sentence format professional engineering instructions;

[0021] Configure the classification output command, requiring the model to classify the generated rework tasks according to the design module or modification type.

[0022] (3) Further, the structured rework task list is stored in JSON format, and each task item includes: task ID, task title, detailed description, priority flag and associated annotation cluster ID index.

[0023] (4) In addition, the present invention also provides an intelligent summary system for collaborative annotation of three-dimensional design based on artificial intelligence, comprising:

[0024] The data acquisition module is configured to acquire the original annotation set generated by the 3D design model during the collaborative review phase.

[0025] The intelligent clustering module is configured to execute a spatial-semantic coupled clustering algorithm to merge discrete annotation objects into annotation clusters; the intelligent clustering module is also configured to perform view frustum overlap detection, and if the volume ratio of the overlapping region is lower than a preset threshold, merging is prohibited.

[0026] The task generation module is configured to use a large language model to summarize and extract the content of the annotation clusters, and generate a structured list of rework tasks.

[0027] The interaction management module is configured to display the rework task list in the interface and realize the viewpoint linkage between the task and the 3D model; the interaction management module is also configured to respond to the user selecting a task in the rework task list, automatically parse the annotation cluster ID associated with the task; calculate a weighted average optimal viewing angle based on the camera viewpoint parameters of all annotation objects contained in the annotation cluster, and control the 3D viewport to smoothly transition to the optimal viewing angle.

[0028] (5) The present invention also provides an electronic device, including a memory, a processor and a computer program stored in the memory and executable on the processor.

[0029] (6) The present invention also provides a computer-readable storage medium having a computer program stored thereon.

[0030] Furthermore, the clustering process employs a variant of DBSCAN (a density-based clustering algorithm), defining a unique distance metric function: Distance = w1 * SpatialDist + w2 * (1 - SemanticSim). Here, SpatialDist is the three-dimensional Euclidean distance, and SemanticSim is the text cosine similarity calculated based on BERT or OpenAI Embeddings. This ensures that only annotations that are "geographically close" and "say the same thing" are merged.

[0031] Furthermore, the task generation step utilizes few-shot learning techniques to guide the large language model to rewrite fragmented spoken comments into a standard engineering instruction format of "verb + noun + parameter".

[0032] The beneficial effects of this invention are as follows:

[0033] 1. Significantly reduces cognitive load. Hundreds or thousands of scattered annotations are automatically compressed into dozens of clear tasks, with a deduplication rate typically exceeding 60%.

[0034] 2. Improve communication accuracy. Utilize AI to transform vague feedback into precise instructions, reducing back-and-forth between designers and reviewers.

[0035] 3. Intelligent viewpoint navigation. The generated tasks come with an "optimal viewing angle," allowing designers to automatically fly to the best editing position by clicking on the task, without the need for manual camera adjustments. Attached Figure Description

[0036] Figure 1 This is a flowchart of the method provided in Embodiment 1 of the present invention;

[0037] Figure 2 This is a schematic diagram illustrating the principle of "spatial-semantic coupling clustering" in an embodiment of the present invention;

[0038] Figure 3 This is a logic diagram of prompt word template construction and task generation in an embodiment of the present invention;

[0039] Figure 4 This is a system architecture block diagram provided in an embodiment of the present invention;

[0040] Figure 5 This is a schematic diagram of the collaborative annotation summary interactive interface provided in an embodiment of the present invention. Detailed Implementation

[0041] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0042] Example 1

[0043] like Figure 1 As shown, the method provided in this embodiment specifically includes the following steps:

[0044] Step S1: Multimodal data acquisition

[0045] The system first extracts all annotation objects for the current version from the collaborative design database.

[0046] The data structure for a single annotation object is defined as follows:

[0047] {

[0048] "annotation_id":"A101",

[0049] "author":"Reviewer_Bob",

[0050] "text":"This chamfer is too sharp and could cut your hand. Make it more rounded."

[0051] "position":{"x":120.5,"y":50.0,"z":-30.2}, / / 3D anchor point

[0052] "camera":{

[0053] "position":{...},

[0054] "target":{...},

[0055] "fov":45

[0056] },

[0057] "timestamp":1678892000

[0058] }

[0059] Step S2: Spatial-Semantic Coupled Clustering

[0060] This is the core algorithm step of the invention. To determine whether two annotations (A and B) are discussing the same issue, the system calculates the comprehensive distance D. ab :

[0061] 1. Spatial proximity calculation: Calculate the Euclidean distance dist(A) between two anchor points. pos B pos If the distance is greater than a preset threshold (e.g., 5% of the model's bounding box diagonal), it is directly determined to be irrelevant (pruning strategy), and semantics are no longer calculated to save computing power.

[0062] 2. View frustum overlap detection (optional): Checks whether the view frustums of cameras A and B are pointing to the same area. This solves the "back-to-back" problem (i.e., two points are very close to each other, but are located on opposite sides of a thin-walled wall, and are actually two unrelated surfaces).

[0063] 3. Semantic Similarity Calculation: A vectorization model (such as text-embedding-ada-002) is used to convert the annotation text into a 1536-dimensional vector, and the cosine similarity Sim is calculated. sem .

[0064] 4. Graph Clustering: Construct an undirected graph, if dist <T spatial And Sim sem >T semantic Then, an edge is established between A and B. Finally, each connected component in the graph is a "comment cluster".

[0065] Step S3: Intelligent Summarization and Task Generation

[0066] For each "annotation cluster", the system concatenates all the original text it contains and fills it into the following prompt word template:

[0067] Prompt template:

[0068] "Character: You are a senior design director."

[0069] Input: The following is a set of original review comments on the same part of the product (which may contain repetitive or emotional content):

[0070] [Comment 1: This is too sharp]

[0071] [Suggestion 2: The chamfer radius has been increased, which is currently unsafe]

[0072] [Opinion 3: Same as above, could easily cause injury]

[0073] Task:

[0074] 1. Understand the core requirements and remove redundant information.

[0075] 2. Rewrite it into a clear, actionable rework task.

[0076] Extract keywords as tags.

[0077] Output format (JSON): {'title':'...','description':'...','tags':[...]}

[0078] Step S4: Result Mapping and Display

[0079] The large language model returns the following structured tasks:

[0080] {

[0081] "task_id":"T001",

[0082] "title":"Optimized Edge Chamfering",

[0083] "description": "The current edge is too sharp and poses a safety hazard. Please increase the chamfer radius to at least 2mm."

[0084] "tags":["security","geometric modification"],

[0085] "source_annotations":["A101","A102","A105"]

[0086] }

[0087] like Figure 5 As shown, the system renders the task card on the right side of the interface. When the user clicks the card, the system calculates a weighted average "optimal viewpoint" based on the camera parameters of the source annotations (A101, A102...) and drives the 3D viewport to smoothly fly into that position.

[0088] Example 2

[0089] This embodiment describes the system architecture design (such as...). Figure 4 (as shown)

[0090] Data acquisition module: responsible for listening to mouse click events on the front-end WebGL canvas and recording anchor point coordinates and raycasting results.

[0091] Intelligent clustering module: Built-in KD-Tree spatial indexing algorithm for accelerating proximity search of large-scale point sets; also integrates vector databases (such as Milvus or Faiss) for semantic retrieval.

[0092] Task generation module: This module encapsulates the large language model API and is responsible for constructing prompt words and parsing results.

[0093] Interaction Management Module: Responsible for rendering the task list, managing its status (pending / completed), and controlling viewpoint animation.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent summarizing collaborative annotations in 3D design based on artificial intelligence, characterized in that, Includes the following steps: Obtain the original annotation set generated during the collaborative review phase of the 3D design model; the original annotation set contains multiple discrete annotation objects, each annotation object containing at least 3D anchor point coordinates, camera view parameters, and unstructured text content; the camera view parameters include viewpoint position, line of sight direction, and field of view angle; The original annotation set is subjected to spatial-semantic coupling clustering to generate several annotation clusters. The clustering process specifically includes: extracting the three-dimensional anchor point coordinates of every two annotation objects, calculating the Euclidean distance, and determining that they are spatially adjacent if the distance is less than a preset spatial threshold; converting the text content of each annotation object into a high-dimensional semantic vector using a word embedding model, calculating the cosine similarity between the vectors, and determining that they are semantically related if the similarity is greater than a preset semantic threshold; performing view frustum overlap detection: determining whether there is an overlapping region between the camera view frustums of two annotation objects, and prohibiting them from being merged into the same annotation cluster even if the spatial distance meets the condition if the overlap region volume ratio is less than a preset threshold; constructing an annotation relationship graph, connecting annotation objects that simultaneously meet the spatial proximity condition, the semantic related condition, and pass the view frustum detection into a connected subgraph, and each connected subgraph constitutes an annotation cluster. Construct a prompt word template that includes role settings and output constraints, serialize all text content and corresponding position information in each annotation cluster and fill it into the prompt word template to generate a large language model input vector; The input vector is fed into a pre-trained large language model for intent understanding and summary generation to obtain a structured list of rework tasks; In the collaborative design interface, the rework task list is associated with the 3D design model and a navigation entry is provided to locate the corresponding annotation cluster space with one click. In response to the user selecting a task in the rework task list, the system automatically parses the annotation cluster ID associated with the task; calculates a weighted average optimal viewing angle based on the camera viewpoint parameters of all annotation objects contained in the annotation cluster; and controls the 3D viewport to smoothly transition to the optimal viewing angle.

2. The method according to claim 1, characterized in that, The constructed prompt word template includes character settings and output constraints, specifically including: Set the model's role as either a design director or an engineering manager; Configure deduplication instructions to instruct the model to identify and merge feedback that contains semantic duplicates; Configure tone conversion instructions to instruct the model to convert colloquial comments into imperative sentence format professional engineering instructions; Configure the classification output command, requiring the model to classify the generated rework tasks according to the design module or modification type.

3. The method according to claim 1, characterized in that, The structured rework task list is stored in JSON format. Each task item includes: task ID, task title, detailed description, priority flag, and associated annotation cluster ID index.

4. An intelligent summary system for collaborative annotation in 3D design based on artificial intelligence, characterized in that: include: The data acquisition module is configured to acquire the original annotation set generated by the 3D design model during the collaborative review phase. The intelligent clustering module is configured to execute a spatial-semantic coupled clustering algorithm to merge discrete annotation objects into annotation clusters; the intelligent clustering module is also configured to perform view frustum overlap detection, and if the volume ratio of the overlapping region is lower than a preset threshold, merging is prohibited. The task generation module is configured to use a large language model to summarize and extract the content of the annotation clusters, and generate a structured list of rework tasks. The interaction management module is configured to display the rework task list in the interface and realize the viewpoint linkage between the task and the 3D model; the interaction management module is also configured to respond to the user selecting a task in the rework task list, automatically parse the annotation cluster ID associated with the task; calculate a weighted average optimal viewing angle based on the camera viewpoint parameters of all annotation objects contained in the annotation cluster, and control the 3D viewport to smoothly transition to the optimal viewing angle.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 3.

Images