Rapid object indexing using locality sensitive hashing and joint 3D-signature space estimation

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• 3D object recognition and indexing are crucial for applications like robotics and augmented reality.
• Existing methods often struggle with real-world data complexity, including occlusions and uncontrolled environments.
• Efficient indexing of large 3D model databases remains a significant challenge.

Purpose of the Study:

• To develop a rapid and robust 3D object indexing method.
• To achieve sublinear complexity for large-scale 3D model databases.
• To maintain high performance with real-world 3D sensed data acquired in uncontrolled settings.

Main Methods:

• Combines feature-based methods with coarse alignment-based matching.
• Utilizes Locality Sensitive Hashing (LSH) for indexing surface descriptors.
• Employs progressively complex geometric constraints to prune initial candidates and eliminate false correspondences.
• Model selection based on the Maximum A Posteriori (MAP) rule using posterior probabilities.

Main Results:

• Demonstrates high performance on real 3D data with occlusions, obscurations, and articulated parts.
• Achieves sublinear complexity relative to the number of models in the database.
• Experiments on a large database (365+ models, 1M features) showed robust vehicle indexing.
• Average processing time per query ranges from 50 to 100 seconds.

Conclusions:

• The proposed method offers an efficient and accurate solution for 3D object indexing in challenging real-world scenarios.
• LSH combined with geometric verification provides a powerful framework for large-scale 3D retrieval.
• The approach effectively handles data acquired in uncontrolled settings, outperforming traditional methods.