Hardware-Efficient Bilateral Filtering for Stereo Matching.

Area of Science:

• Computer Vision
• Image Processing
• Parallel Computing

Background:

• Real-time performance is crucial for practical stereo vision systems, necessitating efficient algorithms.
• Edge-preserving filters, such as the bilateral filter, are effective for high-quality local stereo matching.
• Existing methods may not meet the stringent speed requirements of real-time stereo vision applications.

Purpose of the Study:

• To introduce a novel, hardware-efficient bilateral filtering method tailored for stereo vision.
• To enable high-speed, accurate local stereo matching suitable for real-time processing.
• To demonstrate the filter's utility in cost aggregation for stereo matching algorithms.

Main Methods:

• Development of a hardware-efficient bilateral filter algorithm.
• Implementation and optimization on a Graphics Processing Unit (GPU), specifically an NVIDIA GeForce GTX 580.
• Integration of the filter into a complete local stereo matching pipeline.

Main Results:

• The proposed bilateral filter achieves processing speeds of approximately 417 frames per second for one megapixel color images.
• Quantitative evaluations on the Middlebury benchmark show superior accuracy and speed compared to other local stereo methods.
• The complete stereo matching system, including the filter, achieves an average runtime of 15 milliseconds (67 FPS).

Conclusions:

• The developed hardware-efficient bilateral filter significantly enhances the performance of stereo vision systems.
• This method provides a robust solution for real-time, high-quality stereo matching applications.
• The filter's efficiency and accuracy position it as a valuable component for advanced computer vision tasks.