Object Detection with Improved Shape Context Features

We presented a new approach for detecting objects by combining top-down recognition with improved shape context features and bottom-up segmentation. There are two main steps in this method:hypothesis generation step and verification step. In hypothesis generation step we use top-down recognition with improved shape context features and in verification step we use bottom-up segmentation. Our improved shape context features are more robust to object deformation and background clutter. The shape context is used to generate a set of hypothesis of object location and figure ground mask, which have high recall and low precision rate. In the verification step, we first compute a set of feasible segmentation that are consistent with top-down object hypothesis, then we propose a false positive pruning procedure to prune out false positives. We found the fact that false positive regions typically do not align with any feasible imagesegmentation. From our experiment we show that this simple framework achieve both high recall and high precision with only a few positive training examples and this method can be generalized to many object classes.