Breaking the Rigid Prior: Towards Articulated 3D Anomaly Detection

Problem
Existing 3D anomaly detection methods rely on a rigid prior that assumes normal geometry is pose-invariant, which is inadequate for articulated objects with joints. This limitation leads to misidentification of pose-induced deformations as anomalies, obscuring true structural defects. The authors note that no existing benchmark addresses this challenge, particularly for articulated objects with varying joint configurations. This paper presents ArtiAD, the first large-scale benchmark specifically designed for articulated 3D anomaly detection, comprising 15,229 point clouds across 39 object categories.

Method
The authors introduce a novel approach called Shape-Pose-Aware Signed Distance Field (SPA-SDF). This method replaces the rigid prior with a continuous pose-conditioned implicit field, which is factorized into two components: an articulation-independent structural prior and a Fourier-encoded joint embedding. The training process involves minimizing reconstruction energy to recover the articulation state, allowing for the explicit disentanglement of pose-induced geometry from structural defects. The dataset includes annotations for joint configurations and part-level motion labels, facilitating the evaluation of both interpolation and extrapolation to novel joint configurations. The benchmark also features a seen/unseen articulation split to assess model performance under varying conditions.

Results
SPA-SDF achieves an object-level Area Under the Receiver Operating Characteristic (AUROC) of 0.884 on seen configurations and 0.874 on unseen configurations. These results significantly outperform all rigid-based baselines, demonstrating the effectiveness of the proposed method in accurately detecting anomalies in articulated objects. The paper does not specify the exact performance metrics of the baselines, but the improvement indicates a substantial effect size in the context of articulated anomaly detection.

Limitations
The authors acknowledge that the benchmark is the first of its kind, which may limit comparative studies with existing rigid prior methods. They do not discuss potential scalability issues or the generalizability of the SPA-SDF model to other types of articulated objects not included in the benchmark. Additionally, the reliance on joint configurations for training may restrict the model’s applicability to scenarios with unobserved or novel joint arrangements.

Why it matters
This work addresses a critical gap in 3D anomaly detection for articulated objects, providing a robust benchmark and a novel method that can significantly improve detection accuracy. The introduction of ArtiAD and SPA-SDF has implications for various applications, including robotics, manufacturing, and quality control, where understanding the structural integrity of articulated systems is essential. By facilitating future research in this domain, the authors pave the way for advancements in anomaly detection techniques that can adapt to the complexities of articulated geometries.

Authors: Jinye Gan, Bozhong Zheng, Xiaohao Xu, Junye Ren, Zixuan Zhang, Na Ni, Yingna Wu
Source: arXiv:2604.26868
URL: https://arxiv.org/abs/2604.26868v1