Yuan Tianning ; Wan Fang ; Fu Mengying ; Liu Jianzhuang ; Xu Songcen ; Ji Xiangyang ; Ye Qixiang

Abstract

Despite the substantial progress of active learning for image recognition,there still lacks an instance-level active learning method specified for objectdetection. In this paper, we propose Multiple Instance Active Object Detection(MI-AOD), to select the most informative images for detector training byobserving instance-level uncertainty. MI-AOD defines an instance uncertaintylearning module, which leverages the discrepancy of two adversarial instanceclassifiers trained on the labeled set to predict instance uncertainty of theunlabeled set. MI-AOD treats unlabeled images as instance bags and featureanchors in images as instances, and estimates the image uncertainty byre-weighting instances in a multiple instance learning (MIL) fashion. Iterativeinstance uncertainty learning and re-weighting facilitate suppressing noisyinstances, toward bridging the gap between instance uncertainty and image-leveluncertainty. Experiments validate that MI-AOD sets a solid baseline forinstance-level active learning. On commonly used object detection datasets,MI-AOD outperforms state-of-the-art methods with significant margins,particularly when the labeled sets are small. Code is available athttps://github.com/yuantn/MI-AOD.