Convolutional Channel Features

In this paper, we revisit the multiple channel features approach proposed by Dollar et al., which has shown excellent performances in various computer vision tasks. Enlightened by the ConvNets, we introduce an extended version of multiple channel features called Convolutional Channel Features (CCF), which transfers low-level features from off-the-shelf ConvNet models to feed the boosting classifiers based on decision trees. With the combination of CNN features and decision trees, CCF benefits from the rich capacity, robustness and sparsity in feature representation, as well as more efficiency in computation and storage during inference and learning process. Similar to multiple channel features, CCF is capable to solve diverse vision problems in a sliding window manner, and the computation cost in CCF multi-scale feature pyramid construction can be further reduced with power law based approximation in nearby scales and patchwork for shared convolution. We investigate into a large design space of CCF and show with experiments that CCF achieves leading performances in pedestrian detection, face detection, edge detection and object proposal generation. Codes are available at https://bitbucket.org/binyangderek/ccf.