Dense Out-of-Distribution Detection by Robust Learning on Synthetic Negative Data
This addresses the critical issue of reliable anomaly detection in safety-sensitive applications like autonomous driving and remote sensing, though it is an incremental improvement over existing methods.
The paper tackles the problem of dense out-of-distribution detection in images, where standard models fail on partially anomalous inputs, by generating synthetic negative patches using a normalizing flow and applying an information-theoretic criterion, achieving state-of-the-art results on road-driving and remote sensing benchmarks with minimal computational overhead.
Standard machine learning is unable to accommodate inputs which do not belong to the training distribution. The resulting models often give rise to confident incorrect predictions which may lead to devastating consequences. This problem is especially demanding in the context of dense prediction since input images may be only partially anomalous. Previous work has addressed dense out-of-distribution detection by discriminative training with respect to off-the-shelf negative datasets. However, real negative data are unlikely to cover all modes of the entire visual world. To this end, we extend this approach by generating synthetic negative patches along the border of the inlier manifold. We leverage a jointly trained normalizing flow due to coverage-oriented learning objective and the capability to generate samples at different resolutions. We detect anomalies according to a principled information-theoretic criterion which can be consistently applied through training and inference. The resulting models set the new state of the art on benchmarks for out-of-distribution detection in road-driving scenes and remote sensing imagery, in spite of minimal computational overhead.