Reproduce “Odd-One-Out: Anomaly Detection by Comparing with Neighbors” on OpenResearch

About this paper

Ankan Bhunia, Changjian Li, Hakan Bilen

The research introduces a novel anomaly detection task focused on identifying "odd-looking" objects by comparing instances within a multi-object, multi-view scene, rather than against fixed pre-defined standards. Its "Odd-One-Out" method, which constructs 3D feature volumes and uses sparse voxel attention, achieved 91.78% AUC on the ToysAD-8K-Seen dataset and demonstrated generalization to unseen categories and real-world scenarios.

Key insights

• Anomaly detection can be reframed as a comparative, scene-specific task where "normality" is inferred directly from other instances within the current scene, aligning more closely with human visual inspection processes.
• Distilling knowledge from 2D self-supervised models like DINOv2 into 3D feature volumes significantly enhances part-aware representation and enables robust local correspondences crucial for fine-grained comparison.
• Sparse voxel attention, guided by top-k feature correlations, allows for effective pose-agnostic comparison between object instances by focusing on relevant local feature interactions, reducing noise from irrelevant features.

Summary by alphaXiv.