Research map / Detection and segmentation / Open-set and few-shot detection
Odd-One-Out: Anomaly Detection by Comparing with Neighbors
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces scene-specific multi-object anomaly detection from posed multi-view RGB images, using 3D object-centric feature volumes with DINOv2 feature distillation and sparse voxel cross-instance attention, plus two new synthetic benchmarks (ToysAD-8K, PartsAD-15K).
Problem
Detecting 'odd-looking' objects within a scene by comparing instances to each other, since normality is context-dependent and not fixed by a global standard; occlusions and unknown poses complicate this from single views.
Approach
Builds 3D object-centric feature volumes from sparse multi-view images, enhances them via differentiable rendering with DINOv2 feature distillation for part-aware geometry, and classifies anomalies via sparse voxel attention that matches corresponding regions across instances.
Key results
- Proposed method reaches 91.78 AUC / 83.21 accuracy on ToysAD-8K-Seen, versus 78.13 / 65.55 for DETR3D.
- On ToysAD-8K-Unseen it achieves 89.15 AUC / 81.57 accuracy, and 86.12 AUC / 79.68 accuracy on PartsAD-15K.
- Ablation on ToysAD-8K-Unseen: DINOv2 feature distillation adds 5.8 AUC points and sparse voxel attention adds a further 2.1 AUC over vanilla attention with distillation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Odd-One-Out instead of DETR3D | ToysAD-8K-Seen · ToysAD-8K-Seen | AUC: 79.16 -> 91.78 | Helped |
| use Odd-One-Out instead of DETR3D | ToysAD-8K-Seen · ToysAD-8K-Seen | Accuracy: 67.37 -> 83.21 | Helped |
| use Odd-One-Out instead of DETR3D | ToysAD-8K-Unseen · ToysAD-8K-Unseen | AUC: 74.60 -> 89.15 | Helped |
| use Odd-One-Out instead of DETR3D | ToysAD-8K-Unseen · ToysAD-8K-Unseen | Accuracy: 62.98 -> 81.57 | Helped |
| use Odd-One-Out instead of DETR3D | PartsAD-15K · PartsAD-15K | AUC: 74.49 -> 86.12 | Helped |
| use Odd-One-Out instead of DETR3D | PartsAD-15K · PartsAD-15K | Accuracy: 65.11 -> 79.68 | Helped |
| add vanilla attention across object-centric features (ablation) | Odd-One-Out baseline · ToysAD-8K-Unseen | AUC: 79.13 -> 81.24 | Helped |
| add DINOv2 feature distillation (ablation) | Odd-One-Out with vanilla attention · ToysAD-8K-Unseen | AUC: 81.24 -> 87.05 | Helped |
| add DINOv2 feature distillation (ablation) | Odd-One-Out with vanilla attention · ToysAD-8K-Unseen | Accuracy: 68.20 -> 79.56 | Helped |
| add sparse voxel attention (ablation) | Odd-One-Out with feature distillation · ToysAD-8K-Unseen | AUC: 87.05 -> 89.15 | Helped |
| add sparse voxel attention (ablation) | Odd-One-Out with feature distillation · ToysAD-8K-Unseen | Accuracy: 79.56 -> 81.57 | Helped |
| train on scenes with 3-7 objects and test on 8-12 objects | Odd-One-Out · PartsAD-15K, 8-12 objects | AUC: 86.75 -> 85.10 | Helped |
Surprising findings
- Vanilla attention over object features gives only a slight gain (+2.1 AUC) over a no-comparison baseline, suggesting cross-object attention alone is insufficient without part-aware features.
Limitations
- Targets manufacturing-style anomalies and relies largely on synthetic object shapes rather than real damaged objects.
- Assumes rigid objects and non-touching, not fully occluded instances; no articulation or deformation handling.
- Performance degrades when scenes contain mostly anomalies, and real-world testing is limited to a small set of scenes.
Experimental setup
Datasets: Toys4K ABC dataset ToysAD-8K PartsAD-15K
Benchmarks: ToysAD-8K-Seen ToysAD-8K-Unseen PartsAD-15K
Metrics: AUC accuracy
Baselines: Recon-Recog (COLMAP + DGCNN) ImVoxelNet DETR3D
Builds on: DINOv2 ImVoxelNet DETR3D volume rendering
Closest papers by meaning
- PointAD: Comprehending 3D Anomalies from Points and Pixels for Zero-shot 3D Anomaly Detection (NeurIPS 2024 · similarity 0.87)
- Looking 3D: Anomaly Detection with 2D-3D Alignment (CVPR 2024 · similarity 0.87)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.87)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.86)
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.86)
- Track Any Anomalous Object:A Granular Video Anomaly Detection Pipeline (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.