Area Mapping via Pseudo-Label Discovery for Weakly Supervised Object Detection

Weakly supervised object detection (WSOD) is a task that uses only image-level category labels to train an object detector. The most common weakly supervised object detection framework uses ‘argmax’ as a baseline to filter spatially adjacent pseudo-labels, resulting in poor-quality pseudo-labels. In addition, the pseudo-label screening method based on IoU will lose many potential high-quality labels. To address the above problems, we propose a novel weakly supervised object detection method. We design a pseudo-label benchmark determination method, called area mapping, to determine the pseudo-label benchmark through the feature distribution of image and region proposals and improve the overall quality of pseudo-labels. We further propose a similarity-based pseudo-label discovery strategy utilizing the spatial similarity between pseudo-labels to discover high-quality pseudo-labels. Experiments were conducted on public datasets and new state-of-the-art results were obtained on VOC07 and surpassed the current state-of-the-art methods with a slight advantage on the COCO14 dataset.