Local Context-Aware for Weakly Supervised Object Detection

Weakly supervised object detection (WSOD) is designed to learn the class and location information of objects using image-level labels. Traditional WSOD approaches usually utilize Multiple-Instance Learning (MIL) to create numerous proposals and then choose the most distinctive ones as targets. However, methods that rely on convolutional neural networks tend to emphasize the most salient features of an object, often overlooking the complete object boundaries. In this paper, we propose a single-stage Local Context-Aware (LCA) based Transformer for WSOD. We employ category-wide patch mapping (CWPM) to take into account cross-patch information globally. We also incorporate the Local Awakening Module (LAM), which leverages local features to steer the model’s learning process and enhance weak local responses. Comprehensive experiments on the VOC and COCO datasets showcase the efficacy of LCA. Meanwhile, LCA achieves much faster inference speed than the two-stage WSOD approaches.