Cascade Training as a Tree Search with Dijkstra’s Algorithm

We propose a general algorithm that treats cascade training as a tree search process working according to Dijkstra’s algorithm in contrast to our previous solution based on the branch-and-bound technique. The reason behind the algorithm change is reduction of training time. This change does not affect in anyway the quality of the final classifier. We conduct experiments on cascades trained to become face or letter detectors with Haar-like features or Zernike moments being the input information, respectively. We experiment with different tree sizes and different branching factors. Results confirm that training times of obtained cascades, especially for large heavily branched trees, were reduced. For small trees, the previous technique can sometimes achieve better results but the difference is negligible in most cases.