Online Active Proposal Set Generation for weakly supervised object detection

To reduce the manpower consumption on box-level annotations, many weakly supervised object detection methods which only require image-level have been proposed recently. The training process in these is formulated into two steps. They firstly train a neural network under weak supervision to generate pseudo ground truths (PGTs). Then, PGTs are used another full supervision. Compared with fully methods, becomes more complex and time-consuming. Furthermore, overwhelming negative proposals involved at first step. This neglected by most makes biased towards thus degrades quality of PGTs, limiting performance second Online proposal sampling an intuitive solution issues. However, lacking adequate labeling, simple online may make stuck local minima. solve this problem, we propose Active Proposal Set Generation (OPG) algorithm. Our OPG algorithm consists parts: Dynamic Constraint (DPC) Partition (PP). DPC dynamically determine different strategies according current state. PP score each proposal, part sets active set for optimization. Through experiments, our shows consistent significant improvement both datasets PASCAL VOC 2007 2012, yielding comparable state-of-the-art results.