Test Time Training for Binary Anomaly Segmentation

Deep Learning has recently shown its potential in the Industrial Anomaly Detection and Segmentation context. Nonetheless, there are still problems when it comes to real-world use mainly due to the cold start problem. Also, current methods are particularly good at solving the detection part of anomalies, but not the segmentation one, which is very important in an industrial setting. This is why the main standard technique currently consists in binarizing anomaly score maps. Despite its simplicity, it’s an approach that delivers poor segmentation performance and is subject to too much variability. In this thesis, we propose to apply an innovative approach for training artificial intelligence models at test time, together with a pseudo-labeling mechanism, in order to improve performance on the anomaly segmentation task and mitigate the cold start problem. Our method, named TTT4AS, demonstrates promising results both from a quantitative and qualitative point of view. The finding of this thesis have been published in a CVPR 2024 workshop on anomaly detection.