Dynamic proposal sampling for weakly supervised object detection

Abstract

It is challenging to optimize object detectors with only image-level annotations because the target objects are often surrounded by a large number of background clutters. Many existing approaches tackle this problem through object proposal sampling. However, the collected positive proposals are either low in precision or lack of diversity, and the strategy of collecting negative proposals is not carefully designed, neither. In this context, the primary contribution of this work is to improve weakly supervised detection (WSD) with a dynamic proposal sampling (DPS) strategy. The proposed method collects purified positive training samples by progressively removing confident background clutters, and selects discriminative negative samples by mining class-specific hard proposals. To discover erratic number of confident proposals for different images and categories in varying training phase, we introduce class-specific probabilty accumulation score to measure the image complexity and the quality of learned object detectors, and adjust the number of sampled proposals accordingly. This proposal sampling procedure is integrated into a CNN-based WSD framework, and can be performed in each stochastic gradient descent mini-batch during training. Extensive evaluation results on PASCAL VOC 2007, VOC 2010 and VOC 2012 datasets are presented, which demonstrate that the proposed method effectively improves WSD.