Adding a teaching “assistant”: improving the quality of pseudo-labels for semi-supervised object detection

This paper focuses on semi-supervised object detection (SS-OD) for its tolerance to small amounts of training samples, which is common in real-world applications. Pseudo-label-based approaches have been the mainstream for SS-OD. In this paper, we first show the impact of accurate pseudo-labeling and the challenge of producing such labels. In contrast to prior research that predominantly focused on refining the main model to enhance localization, this paper introduces a novel strategy, where a standalone “Teaching Assistant” or simply “Assistant” is involved in the popular Teacher/Student paradigm to improve the quality of pseudo-labels. This “Assistant” can be plugged into any existing Teacher/Student-based framework without having to fine-tune the original Teacher/Student model. We exploit two “Assistant” models, both of which center around the non-maximum suppression (NMS) method -- a popular technique used to select only the promising bounding boxes. The first “Assistant” model is referred to as the “pre-NMS” assistant that refines the candidate bounding box scores for a better set of inputs to the NMS process. The second “Assistant” model is referred to as the “post-NMS” assistant which takes advantage of SOTA segmentation models to improve the output from the NMS process. We thoroughly evaluate the performance of pre-NMS vs. post-NMS and the impact of improved pseudo-labels on the OD performance. Experimental results on the COCO dataset demonstrate that post-NMS is better than SOTA methods.