Anomaly Segmentation in Foggy Weather for Autonomous Driving with Adaptive Learnable Filters

In autonomous driving, detecting and segmenting out-of-distribution (OOD) samples is crucial. However, this becomes challenging in adverse weather. Existing OOD research mostly focuses on clear weather, leaving a gap for adverse weather OOD data. This paper presents an innovative approach: an adaptive fog parameter predictor and differentiable image dehazing module for foggy anomaly segmentation. The lightweight end-to-end network combines dehazing and anomaly segmentation, leveraging instance normalization, spatial activation functions, and dynamic convolution in dehazing. We enhance Deeplabv3+ for anomaly segmentation and employ Mahalanobis distance-based Standardized Max Logits (SML) for OOD detection and segmentation. Our proposed network undergoes comprehensive comparison with other anomaly segmentation models using the OOD dataset, showcasing its advanced performance in anomaly segmentation of OOD data under foggy scenarios.