DenseHybrid: Hybrid Anomaly Detection for Dense Open-Set Recognition

AbstractAnomaly detection can be conceived either through generative modelling of regular training data or by discriminating with respect to negative data. These two approaches exhibit different failure modes. Consequently, hybrid algorithms present an attractive research goal. Unfortunately, dense anomaly requires translational equivariance and very large input resolutions. requirements disqualify all previous the best our knowledge. We therefore design a novel algorithm based on reinterpreting discriminative logits as logarithm unnormalized joint distribution \(\hat{p}(\textbf{x},\textbf{y})\). Our model builds shared convolutional representation from which we recover three predictions: i) closed-set class posterior \(P(\textbf{y}|\textbf{x})\), ii) dataset \(P(d_{in}|\textbf{x})\), iii) likelihood \(\hat{p}(\textbf{x})\). The latter predictions are trained both standard generic dataset. blend these into score allows open-set recognition natural images. carefully custom loss for in order avoid backpropagation untractable normalizing constant \(Z(\theta )\). Experiments evaluate contributions benchmarks well terms open-mIoU - metric performance. submissions achieve state-of-the-art performance despite neglectable computational overhead over semantic segmentation baseline. Official implementation: https://github.com/matejgrcic/DenseHybridKeywordsDense detectionDense recognitionOut-of-distribution detectionSemantic