STOD: Toward Semi-Supervised Tiny Object Detection

Semi-supervised object detection aims to enhance object detectors by utilizing a large number of unlabeled images, which has gained increasing attention in natural scenes. However, when these methods are directly applied to scenes with tiny objects, they face the challenge of selecting pseudo-labels with high localization quality due to the minuscule and blurred characteristics of these objects. To address this issue, we propose a novel method called semi-supervised tiny object detection (STOD). Firstly, to enhance the localization accuracy of pseudo-labels, we design a dense IoU-aware head that evaluates the quality of bounding box localization by incorporating additional predicted overlap values. Secondly, to mine more potential pseudo-labels, we propose a GMM-based multi-threshold pseudo-labels mining module that dynamically generates multiple thresholds using classification scores and overlap values to classify bounding boxes into strong positive and weak positive pseudo-labels. Lastly, we design the localization-aware weighting loss to incorporate the localization quality of both positive and negative samples in order to enhance the accuracy of pseudo-label localization. The experimental results show that STOD achieves comparable performance when compared to both fully and semi-supervised methods. Notably, on the VisDrone-Partial benchmark, STOD achieves outstanding results by outperforming our baseline model with improvements of 3.1 mAP, 5.7 AP_50 , and 3.0 AP_75 .