Rethinking Pseudo-labeled Sample Mining for Semi-Supervised Object Detection

Duo Li, Sanli Tang, Zhanzhan Cheng, Shiliang Pu, Yi Niu, Wenming Tan, Fei Wu, Xiaokang Yang

Keywords: semi-supervised learning, object detection, sample mining, threshold, consistency

Abstract: Consistency-based method has been proved effective for semi-supervised learning (SSL). However, the impact of the pseudo-labeled samples' quality as well as the mining strategies for high quality training sample have rarely been studied in SSL. An intuitive idea is to select pseudo-labeled training samples by threshold. We find it essential the selection of these thresholds to the final result of SSL. Following this discovery, we propose SEAT (Score Ensemble with Adaptive Threshold), a simple and efficient semi-supervised learning object detection method, in which the high confidence pseudo-labels are selected for self-training. Apart from confidence score as the indicator of the sample's quality, we also introduce the scores of temporal consistency and augmentation consistency. The scores provide a more comprehensive description to the quality of each sample. To cope with the data distribution difference among categories, the adaptive threshold strategy is used to automatically determine the sample mining threshold for each category. We conduct experiments on PASCAL-VOC and MSCOCO, extensive results show that our method is competitive and can be easily combined with consistency-based methods.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

One-sentence Summary: A new method for semi-supervised object detection which can be easily combined with other SOTA methods.

Reviewed Version (pdf): https://openreview.net/references/pdf?id=D-uFMn-EdX