Semi-Supervised Learning for Intelligent Surveillance

Semi-supervised Learning (SSL) has shown promising improvements when used for the Object Detection (OD) task. Existing works detected some unique challenges associated with this setting and discussed how they differ from the classification paradigm. In this work, we revisit two of the first approaches used for OD, namely, Self-Training and Augmentation driven Consistency (STAC) and Unbiased Teacher. Our work was made in partnership with ALTAVE, a Brazilian company focused on intelligent surveillance. By using their in-house dataset we show that relative performance refinement follows the findings from the original authors and propose an improved augmentation method that boosts both approaches, obtaining an overall improvement of 4.4% in the mAP, while using no additional labels for training.