Strategies of Dictionary Usages for Sparse Representations for Pedestrian Classification

Sparse representations and methodologies are currently receiving much interest due to their benefits in image processing and classification tasks. Despite the progress achieved over the last years, there are still many open issues, in particular for applications such as object detection which have been much less addressed from the sparse-representation point of view. This work explores several strategies for dictionary usages and study their relative computational and discriminative values in the binary problem of pedestrian classification. Specifically, we explore whether both class-specific dictionaries are really required, or just any of them can be successfully used and, in case that both dictionaries are required, which is the better way to compute the sparse representation from them. Results reveal that different strategies offer different computational-classification trade-offs, and while dual-dictionary strategies may offer slightly better performance than single-dictionary strategies, one of the most interesting findings is that just one class (even the negative, non-pedestrian class) suffices to train a dictionary to be able to discriminate pedestrian from background images.