Anomaly Pixel Detection via Dual-Branch Uncertainty Metrics

Abstract Anomaly detection aims to detect abnormal samples from normal data. Existing methods mainly use synthesis-based approach and uncertainty-based approach. Moreover, the above-mentioned need re-train part/all of deep network, which limits its application scenarios especially real-time scenarios. We propose a novel retraining-free anomaly method, includes dual-branch mechanism pixels two views. In pseudo-supervised branch, we Copy post method transform unsupervised task into supervised binary task. prototype prototypes improve accuracy pixel detection. The results branches are fused by means soft voting produce with higher recall. Our experiments show that our significantly outperforms other achieves excellent on different datasets, also shows has good generalization ability. Therefore, proposed can effectively effectiveness safety in scenarios, such as autonomous driving.