P2A: Transforming Proposals to Anomaly Masks

Huachao Zhu, Zhichao Sun, Zelong Liu, Yongchao Xu

Abstract: Detecting anomalies in road scenes is essential for safe autonomous driving. Existing methods often consider the likelihood of pixels not belonging to a closed set of classes as the anomaly score. However, this approach lacks object-level understanding and frequently results in numerous false positives at boundaries and ambiguous regions. In this paper, we present a novel method that directly computes the probability of pixels being anomalous and outputs both anomaly segmentation results and score maps. Our approach utilizes the rich semantic information correlated to linguistic concepts in Stable Diffusion to compensate for the low coverage of anomalies caused by limited annotated samples. Using a query-based segmentation model, we transform the proposals into masks of both in-distribution and out-of-distribution objects. Additionally, we introduce an image-mask-image pipeline to generate various annotated data as outliers for supervised training. Extensive experiments across multiple benchmarks confirm that the proposed method outperforms previous state-of-the-art methods in road anomaly segmentation. Code is available at https://github.com/huachao0124/P2A.