PTAD: Pose and Texture Agnostic Anomaly Detection

Wei Zhuo, Jianen Xiang, Miaomiao Liu, Huajun Lu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings, 2026, pp. 6779-6788

Anomaly detection based on images is a core task in industrial inspection and defect recognition. Recent methods perform well in controlled environments with pose alignment and texture consistency. However, when faced with real-world conditions where poses and surface textures vary, existing methods degrade significantly. This paper introduces the pose and texture-agnostic anomaly detection problem, which generalizes the pose-unknown setting by using a texture-free 3D reference as a geometric prior to detect and localize structural anomalies from arbitrary viewpoints. Unlike training feature embedding based on large-scale multi-view images, we propose a novel render-for-detect framework that targets to a challenging setting of inspecting geometric anomalies depending on texture-less references. Specifically, we first reconstruct a 3DGS from texture-less references in anomaly detection and render the target pose image from mask-guided pose estimation with differential depth rendering. Unlike prior single-modal detection, an efficient multi-modal anomaly detection network is introduced, that is proven boundary-sensitive. Benefiting from the general embedding, we can gain competitive results with boundary feature improvements and only using 20% training data, making the framework more feasible in real application scenarios.