PANDA: Adapting Pretrained Features for Anomaly Detection and Segmentation

Anomaly detection methods require high-quality features. In recent years, the\nanomaly detection community has attempted to obtain better features using\nadvances in deep self-supervised feature learning. Surprisingly, a very\npromising direction, using pretrained deep features, has been mostly\noverlooked. In this paper, we first empirically establish the perhaps expected,\nbut unreported result, that combining pretrained features with simple anomaly\ndetection and segmentation methods convincingly outperforms, much more complex,\nstate-of-the-art methods.\n In order to obtain further performance gains in anomaly detection, we adapt\npretrained features to the target distribution. Although transfer learning\nmethods are well established in multi-class classification problems, the\none-class classification (OCC) setting is not as well explored. It turns out\nthat naive adaptation methods, which typically work well in supervised\nlearning, often result in catastrophic collapse (feature deterioration) and\nreduce performance in OCC settings. A popular OCC method, DeepSVDD, advocates\nusing specialized architectures, but this limits the adaptation performance\ngain. We propose two methods for combating collapse: i) a variant of early\nstopping that dynamically learns the stopping iteration ii) elastic\nregularization inspired by continual learning. Our method, PANDA, outperforms\nthe state-of-the-art in the OCC, outlier exposure and anomaly segmentation\nsettings by large margins.\n