Object Proposal By Hierarchical Segmentation

Object proposal method has become an important preprocessing step in modern object detection paradigm,and has been successfully applied in other computer vision problems,such as video-based object detection and tracking,3D object recognition.Object detection system uses object proposal method to reduce the number of image windows for detection from millions to just a few hundreds.This greatly reduces the time cost for real-world object detection systems,and enables them to use far more complicated classifiers,such as a deep convolutional neural network with nearly 100 layers.In this thesis,we propose our novel object proposal method: multi-branched hierarchical segmentation based on a general principle: an object consists of subparts with various colors and textures may require a combination of different merging strategies to bring it together.Hence,we propose to try different similarity measures throughout the bottom-up merging process to search the segment composition space more effectively.To make our multi-branched hierarchical segmentation effective,we propose an automatic learning procedure that trains the complementary similarity measures by altering weights for each training samples in a Adaboostlike fashion.Different from standard AdaBoost,our goal is not to combine the classifiers into a strong one,but to obtain a set of complementary ones such that errors in one similarity measure could be corrected by the others.Extensive comparisons to previous object proposal methods indicate that our approach achieves the state-of-the-art results in terms of object proposal evaluation protocols.Moreover,in order to investigate how well these methods perform for real-world object detection tasks,we test all compared object proposal methods using the state-of-the-art R-CNN detector on PASCAL VOC2007 test set.As a result,our approach achieves the best mAP rate with 30%less window proposals than Selective Search.