Research map / Detection and segmentation / Open-set and few-shot detection
PIAD: Pose and Illumination agnostic Anomaly Detection
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 Pose and Illumination agnostic Anomaly Detection (PIAD), a new dataset with synthetic and real industrial objects, and a 3DGS-based baseline that uses reflectance maps from intrinsic decomposition for pose optimization and anomaly scoring.
Problem
Anomaly detection when query images differ from training images in both camera pose and illumination, which breaks pose-agnostic methods that bake lighting into radiance fields.
Approach
Train a 3D Gaussian Splatting model on both RGB and illumination-invariant reflectance maps (from Retinex-based intrinsic decomposition), register queries by optimizing camera pose with a combined reflectance and color L1 loss, and score anomalies by comparing multi-scale CNN features of color and reflectance.
Key results
- On the PIAD real dataset, the method reaches mean pixel AUROC 99.05 and image AUROC 92.63, versus 94.86/73.20 for OmniposeAD and 98.01/79.82 for SplatPose.
- On MAD-Sim the method achieves image AUROC 99.5 pixel-level vs 99.0 for SplatPose; on the PIAD synthetic set image AUROC 96.7 vs 84.9 for OmniposeAD.
- Pose estimation under PIAD gives AUC@5° of 65.22 vs 64.11 for SplatPose and 27.36 for OmniposeAD; anomaly detection takes 7.42 s per object vs 51.70 s for OmniposeAD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PIAD instead of SplatPose | Anomaly detection · MAD-Sim | Pixel AUROC: 99.0 -> 99.5 | Helped |
| Use PIAD instead of SplatPose | Anomaly detection · MAD-Sim | Image AUROC: 94.9 -> 97.4 | Helped |
| Use PIAD instead of SplatPose | Anomaly detection · Our dataset (synt) | Pixel AUROC: 97.4 -> 99.0 | Helped |
| Use PIAD instead of SplatPose | Anomaly detection · Our dataset (synt) | Image AUROC: 85.9 -> 96.7 | Helped |
| Use PIAD instead of SplatPose | Anomaly detection · Our dataset (real) | Pixel AUROC: 98.01 -> 99.05 | Helped |
| Use PIAD instead of SplatPose | Anomaly detection · Our dataset (real) | Image AUROC: 79.82 -> 92.63 | Helped |
| Use reflectance initialization and color-plus-reflectance optimization instead of color initialization and optimization (ablation) | PIAD pose estimation · Teapot and Wand | Image AUROC: 87.75 -> 91.10 | Helped |
| Use color and reflectance for pose optimization instead of reflectance-only optimization (ablation) | PIAD pose estimation · Teapot and Wand | Pixel AUROC: 97.55 -> 98.60 | Helped |
| Set pose-loss reflectance weight λ to 0.6 rather than 0 (ablation) | PIAD pose optimization · Teapot and Wand | Image AUROC: 88.0 -> 91.1 | Helped |
| Set pose-loss reflectance weight λ to 0.6 rather than 0.9 (ablation) | PIAD pose optimization · Teapot and Wand | Image AUROC: 87.1 -> 91.1 | Helped |
| Set pose-loss reflectance weight λ to 0.6 rather than 1.0 (ablation) | PIAD pose optimization · Teapot and Wand | Image AUROC: 88.4 -> 91.1 | Helped |
| Include color and reflectance features together instead of color features alone (ablation) | PIAD anomaly scoring · Teapot and Wand | Image AUROC: 89.80 -> 91.10 | Helped |
Surprising findings
- Reflectance-based pose initialization and optimization improves image AUROC notably on the real dataset, while color features are more accurate at pixel-level detection.
- Table 6 shows the pose-estimation gains for the proposed method over SplatPose are small at AUC@5° (65.22 vs 64.11) even though the translation AUC gains are large.
Limitations
- Pose-estimation AUC at 5° under PIAD is only modestly better than SplatPose.
- Pixel AUROC drops for the proposed method under some illumination directions compared with baselines.
- Reliance on a pre-trained intrinsic decomposition network and a pre-trained matcher for initialization.
Experimental setup
Datasets: MAD-Sim PIAD dataset (synthetic) PIAD dataset (real)
Benchmarks: MAD-Sim PIAD dataset (synt) PIAD dataset (real)
Metrics: pixel AUROC image AUROC pose estimation AUC runtime
Baselines: OmniposeAD SplatPose
Builds on: 3D Gaussian Splatting OmniposeAD SplatPose URetinexNet EfficientLoFTR
Closest papers by meaning
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.84)
- AnomalyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection (ICLR 2024 · similarity 0.84)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.84)
- PointAD: Comprehending 3D Anomalies from Points and Pixels for Zero-shot 3D Anomaly Detection (NeurIPS 2024 · similarity 0.83)
- Track Any Anomalous Object:A Granular Video Anomaly Detection Pipeline (CVPR 2025 · similarity 0.83)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.83)
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.