P2W: Improving the Cognition of Object Wholeness for Weakly Supervised Object Detection
Lei Fang, Jin Dang, Xiaofen Tang · 2023
In the Weakly Supervised Object Detection (WSOD) task, the detector is trained solely by image-level labels, which significantly reduced the learning cost. However, in doing so, the model lacks precise instance-level supervision, which hinders its ability to accurately understand the full object structure. As a result, the model tends to focus on easily distinguishable local regions and make imprecise predictions. To address this problem, this paper proposed a Part to Whole (P2W) approach, consisting of two main components: the Hierarchical Relationship Extraction (HRE) module and the Object Component Integration (OCI) module. The HRE module synchronously extracts hierarchical relationships among proposals in conjunction with Selective Search. The OCI module uses these relations as clues to explore higher-level proposals for seeking core boxes and generating more complete object bounding boxes. With these bounding boxes as pseudo-labels, the model's understanding of object wholeness is enhanced and hence its performance will be better. Experimental result on the VOC 2007 reached 59.3% mAP, becoming the new state-of-the-art result. And the result on VOC 2012 reached 54.3% mAP, also closing to the state-of-the-art result.