Far Away in the Deep Space: Nearest-Neighbor-Based Dense Out-of-Distribution Detection
share
The key to out-of-distribution detection is density estimation of the in-distribution data or of its feature representations. While good parametric solutions to this problem exist for well curated classification data, these are less suitable for complex domains, such as semantic segmentation. In this paper, we show that a k-Nearest-Neighbors approach can achieve surprisingly good results with small reference datasets and runtimes, and be robust with respect to hyperparameters, such as the number of neighbors and the choice of the support set size. Moreover, we show that it combines well with anomaly scores from standard parametric approaches, and we find that transformer features are particularly well suited to detect novel objects in combination with k-Nearest-Neighbors. Ultimately, the approach is simple and non-invasive, i.e., it does not affect the primary segmentation performance, avoids training on examples of anomalies, and achieves state-of-the-art results on the common benchmarks with +23 StreetHazards respectively.
READ FULL TEXT