Dynamic updating self-training for semi-weakly supervised object detection

Abstract

Semi-weakly supervised object detection (SWSOD) needs to label only a small portion of images in the training set to train an initial detector. This detector is then used to select some reliable unlabeled images and generate pseudo labels for them. The pseudo-labeled images are combined with the labeled images to re-train the detector. One potential problem of SWSOD is that some pseudo labels could be incorrect so that the detector cannot be improved progressively. To address this problem, we propose a dynamic updating self-training (DUST) mechanism, which divides unlabeled images into multiple folds ranging from simple to complex and dynamically updates pseudo labels. In each fold, we apply an intra-fold updating at each training iteration to iteratively select some images from the current fold and update their pseudo labels to retrain the detector. At the end of each fold, we apply an inter-fold updating to update the pseudo labels of all previous folds and transfer the remaining unlabeled images into the next fold for further pseudo labels mining. To reduce the influence of noisy labels during the intra-fold updating, we propose a label reliability sensitive (LRS) loss for the labeled images to weigh the cross entropy. Meanwhile, a sample difficulty sensitive (SDS) loss is proposed for unlabeled images to balance the contributions of samples with different confidences. Extensive experiments on PASCAL VOC and MS-COCO benchmarks demonstrate the effectiveness of our method. For instance, by using 20 positive images and 20 negative images in each category as the labeled images, our method outperforms the state-of-the-art method [46] by 4.1% and achieves 93% of the WSOD performance on the PASCAL VOC 2007 dataset.