EviOcc: Preserving Visual Evidence for Cross-Dataset 3D Occupancy Anomaly Detection

Autonomous vehicles need to locate unexpected obstacles even when their categories are absent from training data. Occupancy anomaly prediction addresses this need by combining scene geometry and semantics with voxel-level anomaly detection. However, the image features used to detect unfamiliar objects also change with camera geometry and scene appearance. When a model is transferred across datasets, these changes can cause normal objects to be flagged as anomalies. Addressing these errors through additional data collection, annotation, and model adaptation for each new camera configuration or environment is costly. We therefore introduce EviOcc, a source-only framework that predicts semantic occupancy and anomaly scores from a single RGB image and camera calibration. Using matching patches in resized and cropped versions of the same source image, EviOcc learns normal features expected at each patch's distance, viewing direction, and sampling scale. Deviations from these expected features provide anomaly scores for individual patches before feature pooling. Estimated depth locates each patch in 3D, where local kernels predict semantic occupancy and distribute patch anomaly scores to voxels. Directly projected scores are also retained, even at voxels predicted as free. Experiments across four training sources and three anomaly benchmarks demonstrate strong same-family and cross-dataset anomaly detection without target-domain adaptation, together with competitive geometric and semantic occupancy accuracy on source-domain and same-family evaluations.