Insatiate boosted forest: Towards data exploitation in object detection

Farzin Ghorban, Yu Su, Mirko Meuter, Anton Kummert

Abstract: Boosted forest (BF) is a commonly used method for object detection. With the help of cascade strategy, it can efficiently reject non-object windows and finally, combined with sliding window paradigm, give the locations of target objects in an image. In the literature, many aspects of cascaded boosted forest (CBF) have been well studied, such as image representation, tree split and cascade structure. Although it has been extensively investigated, CBF is still not saturated. In this work, we demonstrate by a series of experiments that the performance of a CBF-based object detector can be significantly improved by careful data exploitation. Specifically, we use a simple yet efficient approach to collect more training samples with high quality. We show that the trained CBF-based detector can be included in the data collection loop to provide better training samples. Our experiments are conducted on two challenging pedestrian detection benchmarks: Caltech and Kitti. Taking Caltech for example, we get the best balance between performance and run-time: 17.2% miss rate (MR) while running at 11 frames per second (FPS) on a moderate CPU. Compared with state-of-the-art CBF-based detectors our method gives better or similar performance while running significantly faster.