Weakly supervised foreground learning for weakly supervised localization and detection

Modern deep learning models require large amounts of accurately annotated data, which is often difficult to satisfy. Hence, weakly supervised tasks, including object localization (WSOL) and detection (WSOD), have recently received attention in the computer vision community. In this paper, we motivate propose foreground (WSFL) task by showing that both WSOL WSOD can be greatly improved if groundtruth masks are available. More importantly, a complete WSFL pipeline with low computational cost, generates pseudo boxes, learns masks, does not need any annotations. With help predicted our model, achieve 74.37% correct accuracy on CUB for WSOL, 55.7% mean average precision VOC07 WSOD, thereby establish new state-of-the-art tasks. Our model also shows excellent transfer ability.