Detecting Binocular Half-Occlusions: Empirical Comparisons of Five Approaches
IEEE Transactions on Pattern Analysis and Machine Intelligence
Image-Gradient-Guided Real-Time Stereo on Graphics Hardware
3DIM '05 Proceedings of the Fifth International Conference on 3-D Digital Imaging and Modeling
Symmetric Stereo Matching for Occlusion Handling
CVPR '05 Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05) - Volume 2 - Volume 02
High-Quality Real-Time Stereo Using Adaptive Cost Aggregation and Dynamic Programming
3DPVT '06 Proceedings of the Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06)
Stereo Processing by Semiglobal Matching and Mutual Information
IEEE Transactions on Pattern Analysis and Machine Intelligence
IEEE Transactions on Pattern Analysis and Machine Intelligence
A fast and parallel belief computation structure for stereo matching
EurolMSA '07 Proceedings of the Third IASTED European Conference on Internet and Multimedia Systems and Applications
Real-Time System for High-Image Resolution Disparity Estimation
IEEE Transactions on Image Processing
| Hi-index | 0.00 |
High-definition 3D video is one of the features that the next generation of telecommunication systems is exploring. Real-time requirements limit the execution time of stereo-vision techniques to 40-60 milliseconds. Classical belief propagation algorithms (BP) generate high quality depth maps. However, the huge number of required memory accesses limits their applicability in real systems. This paper proposes a real-time (latency inferior to 40 millisenconds) high-definition (1280x720) stereo matching algorithm using belief propagation. There are two main contributions. The first one is an improved BP algorithm with occlusion, potential errors and texture-less handling that outperforms classical multi-grid bipartite-graph BP while reducing the number of memory accesses. The second one is an adaptive message compression technique with low performance penalty that greatly reduces the memory traffic. The combination of these techniques outperforms classical BP by about 6.0% while reducing the memory traffic by more than 90%.