Semi-Supervised Learning with Dense Target Producer for End-to-End Lightweight Polyp Detection

Semi-Supervised Object Detection (SSOD) can greatly enhance the performance of object detectors by leveraging a vast amount of unlabeled data. Despite great successes, most existing SSOD methods are based on pseudo-boxes, which require subsequent postprocessing steps with many hyperparameters to be tuned. This paper proposes a novel effective semi-supervised method called Dense Target Producer (DTP) for heatmap-based end-to-end detectors. Unlike conventional SSOD methods, our method directly uses dense prediction of a teacher model as pseudo-labels without the need for any postprocessing or target assigner steps. Moreover, we also introduce Threshold Epoch Adaptor (TEA), a dynamic thresholding procedure to adaptively filter unreliable pseudo labels based on the learning status of the models. Finally, we propose a lightweight heatmap-based end-to-end detector, namely CenterNet++, that enhances the performance relative to the baseline CenterNet. Experiments conducted on a large PolypsSet dataset across different settings showcase our method's superior performance over existing SSOD methods. The experimental results show that our DTP significantly increases the AP50 score by 16.7, 3.2, and 2.8 compared to the supervised baseline model when using 1%, 5%, and 10% of the training set as labeled data, respectively.