Parallel Structured Mesh Generation with Disparity Maps by GPU Implementation.

Area of Science:

• Computer Vision and Graphics
• Computational Geometry
• Machine Learning

Background:

• Generating detailed 3D surface representations is crucial for various applications.
• Existing methods often struggle with scalability and real-time performance for complex scenes.
• Disparity maps provide depth information but require efficient processing for detailed surface reconstruction.

Purpose of the Study:

• To develop a compressed 3D surface representation prioritizing detail for near objects based on disparity maps.
• To leverage the Kohonen's Self-Organizing Map algorithm for efficient and parallelizable 3D surface compression.
• To achieve near real-time performance for large-scale 3D reconstruction tasks.

Main Methods:

• Utilized Kohonen's Self-Organizing Map (SOM) algorithm for generating a topological map from disparity data.
• Partitioned the disparity map into cell units, each associated with a processing unit for decentralized data decomposition.
• Implemented a GPU-accelerated solution for massively parallel processing.

Main Results:

• The proposed method generates a compressed 3D surface representation with detail proportional to object proximity.
• Experimental results demonstrate near real-time performance on GPU, with linear scaling of running time.
• The approach exhibits a linearly increasing relationship between processing units/memory and problem size.

Conclusions:

• The decentralized, data-decomposition approach based on Kohonen's SOM is effective for 3D surface compression.
• The GPU implementation offers efficient, scalable performance suitable for large-scale 3D reconstruction.
• This method provides a robust solution for generating detailed 3D models in near real-time.