Object proposal generation has become a vital step in many vision tasks, like object detection and object instance segmentation. While great progress has been made in this area, there still are several serious limitations in the prior methods. First, most object proposal approaches focus on 2D images and are unable to make use of multi-modal cues. Second, generating accurate object segment proposals is still an extremely challenging task. Finally, the refinement of object proposals is rarely investigated. In our work, we focus on developing algorithms to produce high-quality object proposals, especially the segment proposals. To this end, we first extend the object proposal generation to stereo images and explore the geometric information as well as semantic context in object proposal generation. Second, we introduce the representation learning and similarity learning into the process of grouping superpixels into meaningful objects, to produce high-quality segment proposals. Finally, we propose to improve the precision of object proposals through warping them by learning spatial transformers. We evaluate all our methods on several publicly available object recognition datasets and show the effectiveness of our approaches.
Haoyang Zhang is a PhD student in ANU, supervised by Xuming He, Laurent Kneip, Lexing Xie and Miaomiao Liu. His current research field is object proposal generation.