Learning for stereo vision using the structured support vector machine

We present a random field based model for stereo vision with explicit occlusion labeling in a probabilistic framework. The model employs non-parametric cost functions that can be learnt automatically using the structured support vector machine. The learning algorithm enables the training of models that are steered towards optimizing for a particular desired loss function, such as the metric used to evaluate the quality of the stereo labeling. Experimental results demonstrate that the performance of our method surpasses that of previous learning approaches and is comparable to the state-of-the-art for pixel-based stereo. Moreover, our method achieves good results even when trained on different image sets, in contrast with the common practice of hand tuning to specific benchmark images. In addition, we investigate the impact of graph structure on model performance. Our study shows that random field models with longer-range edges generally outperform the 4-connected grid and that this advantage is especially pronounced for noisy images.