Statistical model for intensity differences of corresponding points between stereo image pairs

Correspondence analysis is required for many applications such as multimedia communication and 3-D telepresence. Current intensity-based approaches model intensity differences of corresponding points in the left- and right-eye images with Gaussian distribution. In this contribution, the statistical characteristics of intensity differences of corresponding points were studied using natural stereo images. The examination reveals that a Laplacian distribution outperforms a Gaussian distribution. Based on this result, a new approach for correspondence analysis is proposed, which exploits Laplacian distribution to model intensity differences of corresponding points. To measure the performance of different approaches, a measure related to the peak signal-to-noise ratio (PSNR) of disparity-compensated prediction over the matching ratio was introduced. The experimental results show that the proposed correspondence algorithm has a better performance than other existing approaches. It also shows that the PSNR of disparity-compensated prediction decreases as the matching ratio goes up.