Context augmentation for object detection

Current two-stage object detectors, which mainly consist of a region proposal stage and a proposal recognition stage, may produce unreliable results for objects appearing with little information such as small and occluded objects. This is caused by poor region proposals and inaccurate proposal recognition. To address this problem, we propose a context augmentation algorithm that fully utilizes contextual information to generate high-quality region proposals and detection results. First, Region proposals are produced by two steps: 1) generate a coarse set of region proposals, some of which are reliable and some of which are ambiguous, and 2) the ambiguous region proposals are re-estimated using appearance and geometry information with respect to the reliable region proposals from step 1). Second, similar types of pair-wise relations between region proposals are used to produce global feature information associated with the region proposals in order to enhance recognition results. In practice, our method effectively improves the quality of region proposals as well as recognition results. Empirical studies show that the proposed context augmentation yields substantial and consistent improvements over baseline Faster R-CNN. Moreover, there is around 1.3% mAP improvement over Mask R-CNN on COCO dataset.