RAD: A Dataset and Benchmark for Real-Life Anomaly Detection with Robotic Observations
RAD is a robot-captured benchmark for detecting anomalies in everyday objects under uncontrolled conditions. It combines 5,848 images from various viewpoints with pixel-level defect annotations. The research shows that mature 2D feature methods outperform 3D and vision-language models, highlighting the need for models that handle appearance, geometry, and viewpoint uncertainty jointly.