Comprehensive Criterion-Guided Variable Memory Bank for Semi-Supervised Object Detection
Yaokun Yang, Ziyuan Yang, Peng Wei Wang, Yi Zhang · 2024
Semi-supervised object detection methods (SSOD) based on self-training heavily rely on pseudo labels derived from unlabeled data, but these methods often require manual thresholds to remove poor pseudo boxes. In addition, there’s a lack of reliable indicators for both location and classification validation. To relieve the above problems, we propose a novel indicator for SSOD, dubbed as COmprehensive criterion-guided variable Memory Bank (COMB). Specifically, COMB is a comprehensive criterion designed to simultaneously assess the location and classification accuracy. The proposed criterion uniquely maps the two terms into Euclidean space, maintaining their independence. To avoid introducing manual prior knowledge, COMB utilizes a clustering method with the Jensen-Shannon divergence to gauge the distances between pseudo boxes through the teacher model. Then, the optimal pseudo boxes are selected for inclusion in the variable memory bank, which is used for training the student model. Extensive experiments are conducted to validate the effectiveness of the proposed COMB in comparison with state-of-the-art approaches on the public benchmark datasets, MS-COCO and PASCAL-VOC.