Active Exploration of Large 3D Model Repositories

Area of Science:

• Computer Science
• Information Retrieval
• Computer Vision

Background:

• Large-scale 3D model repositories necessitate efficient exploration methods.
• Existing retrieval techniques struggle with scalability and limited search refinement options.

Purpose of the Study:

• To develop an interactive approach for efficient exploration of large 3D model repositories.
• To enhance 3D model retrieval by incorporating user feedback and inter-model similarities.

Main Methods:

• An active learning procedure is employed, driven by user labeling of models or parts as 'like' or 'dislike'.
• Offline pre-processing computes global and local shape descriptors for each model.
• A sparse distance metric is derived for efficient evaluation on large databases, considering both query-model and model-model similarities.

Main Results:

• The proposed method enables efficient interactive exploration of a 3D model repository exceeding 100,000 models.
• User feedback effectively refines search results, overcoming limitations of traditional retrieval methods.
• The system presents candidate models based on estimated relevance, improving user experience.

Conclusions:

• The interactive approach significantly improves the efficiency and effectiveness of 3D model retrieval from large-scale repositories.
• Exploiting similarities among database models alongside query-model similarity is key to scalability.
• This method offers a robust solution for navigating and discovering relevant 3D models in extensive collections.