Distilling Knowledge from Refinement in Multiple Instance Detection Networks

Weakly supervised object detection (WSOD) aims to tackle the object detection\nproblem using only labeled image categories as supervision. A common approach\nused in WSOD to deal with the lack of localization information is Multiple\nInstance Learning, and in recent years methods started adopting Multiple\nInstance Detection Networks (MIDN), which allows training in an end-to-end\nfashion. In general, these methods work by selecting the best instance from a\npool of candidates and then aggregating other instances based on similarity. In\nthis work, we claim that carefully selecting the aggregation criteria can\nconsiderably improve the accuracy of the learned detector. We start by\nproposing an additional refinement step to an existing approach (OICR), which\nwe call refinement knowledge distillation. Then, we present an adaptive\nsupervision aggregation function that dynamically changes the aggregation\ncriteria for selecting boxes related to one of the ground-truth classes,\nbackground, or even ignored during the generation of each refinement module\nsupervision. Experiments in Pascal VOC 2007 demonstrate that our Knowledge\nDistillation and smooth aggregation function significantly improves the\nperformance of OICR in the weakly supervised object detection and weakly\nsupervised object localization tasks. These improvements make the Boosted-OICR\ncompetitive again versus other state-of-the-art approaches.\n