In a dynamic, programming-based stereopsis algorithm, the matching algorithm can be described as the process of constructing a continuous path through possible left-right image location matches. Choosing the best path involves the selection of an appropriate left-right matching function, as well as an appropriate regularization process to handle regions that are not visible binocularly, and to handle failures of the left-right matching function. This paper presents an overview of the elements of dynamic programming-based stereopsis algorithms, and provides an analysis of the effects of different design choices within the algorithm on its overall performance.
The paper presents an in-depth analysis of the effects of two design choices: the left-right similarity function and regularization. Two implementations are described. In the first, the matching process uses a more traditional algorithm, within which the similarity function and regularization process are represented as two different processes. In the second, the matching process is described as a Markov model, with transition probabilities controlled by the both the similarity and regularization terms.
Results presented with both algorithms demonstrate that no global “best” similarity function-regularization process exists (at least, for the class of algorithm and similarity/regularization approaches presented here). The paper suggests that a local process to locally tune the regularization within a given image would be a promising approach for future research. I found this conclusion to be somewhat dissatisfying. Experiments with other stereo algorithms have demonstrated that best matching functions and regularization processes are elusive, and that approaches that work well for one class of images may or may not work well for another. Locally adapting the regularization/matching process is certainly an interesting direction for further research, and it would have been more satisfying if the paper had proposed such directions more explicitly, at least for the class of dynamic programming-based stereopsis algorithms.
Although the paper is highly motivated by the author’s earlier work in this area, it does provide a balanced review of the work in dynamic programming-based stereopsis algorithms. The paper would be suitable for stereovision researchers with interests in dynamic programming-based approaches, and would be a good starting point for graduate students and researchers who are contemplating research in the area.