Refining and Selecting Pseudo Ground Truth for Weakly-Supervised Object Detection

Abstract

Weakly supervised object detection (WSOD) is an important issue in vision tasks. Unlike fully supervised learning, weakly supervised learning uses only image-level labels without bounding boxes. Training with only image-level labels makes it difficult to train deep networks based detectors in a weakly supervised manner. This paper proposes methods which are denoted as refining pseudo ground truth (RPG) and selecting pseudo ground truth (SPG), respectively. Pseudo ground truth for 1st instance classifier refinement network is refined to make better bounding boxes with RPG and select good bounding boxes with SPG. The proposed methods obtain 55.75% as a mean average precision (mAP) on VOC 2007 that outperforms the previous state-of-the-art.