Research map / Detection and segmentation / 3D detection and driving
Look Inside for More: Internal Spatial Modality Perception for 3D 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.
Proposes ISMP, a 3D point-cloud anomaly detection method that adds internal spatial pseudo-modality projections (Spatial Insight Engine), FPFH-enhanced key-point features, and a Laplacian feature filtering module, evaluated on Real3D-AD and Anomaly-ShapeNet.
Problem
3D anomaly detection methods focus on external structure of point clouds and underuse internal structural information, limiting localization and detection accuracy.
Approach
Projects point clouds into internal 2D pseudo-modal slices (top/middle/bottom, top-down and bottom-up) via a Spatial Insight Engine to capture global internal information beyond external projections, combined with FPFH-enhanced PointMAE patch features and a Laplacian-based feature filter feeding dual feature and coordinate memory banks.
Key results
- ISMP reaches mean O-AUROC 0.757 and P-AUROC 0.836 on Real3D-AD, reported as 3.2% object-level and 13.1% pixel-level improvements over prior SOTA.
- On Anomaly-ShapeNet, ISMP achieves mean P-AUROC 0.691, the best among compared methods.
- PointNet++ with SIE features improves ModelNet40 accuracy from 90.7% to 91.1% and ShapeNet-Part IoU from 85.1% to 85.4%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the full ISMP method | Real3D-AD anomaly detection · Real3D-AD | O-AUROC: 0.725 -> 0.767 | Helped |
| add the full ISMP method | Real3D-AD anomaly detection · Real3D-AD | P-AUROC: 0.733 -> 0.836 | Helped |
| add the full ISMP method | Anomaly-ShapeNet anomaly detection · Anomaly-ShapeNet | P-AUROC: 0.668 -> 0.691 | Helped |
| add SIE (ablation) | ISMP · Real3D-AD | I-AUROC: 0.656 -> 0.757 | Helped |
| add SIE (ablation) | ISMP · Real3D-AD | P-AUROC: 0.827 -> 0.836 | Helped |
| add feature filtering (ablation) | ISMP without SIE · Real3D-AD | P-AUROC: 0.809 -> 0.827 | Helped |
| add enhanced feature extraction (ablation) | ISMP without SIE and feature filtering · Real3D-AD | P-AUROC: 0.620 -> 0.809 | Helped |
| add internal projection slices (ablation) | ISMP using external projection only · Real3D-AD | I-AUROC: 0.717 -> 0.757 | Helped |
| add internal projection slices (ablation) | ISMP using external projection only · Real3D-AD | P-AUROC: 0.812 -> 0.836 | Helped |
| add SIE features | PointNet++ · ModelNet40 | Accuracy: 90.7 -> 91.1 | Helped |
| add SIE features | PointNet++ · ShapeNet-Part | IoU: 85.1 -> 85.4 | Helped |
| use the full ISMP model rather than RegAD | 3D anomaly detection on seahorse class · seahorse class | Training FPS: 5.52 -> 2.78 | Hurt |
Surprising findings
- Using only the two external projection slices (ISMP w/o internal slice) gives the second-best results, and the table shows I-AUROC dropping from 0.757 to 0.717 without internal projection, indicating the internal slice contributes but external projections alone remain strong.
Limitations
- Inference and training speed lag behind some competitors due to extra modality information, as acknowledged by the authors.
- Feature filtering parameters (alpha, beta, gamma) were chosen conservatively from heatmaps rather than systematically tuned.
- Theoretical justification of information gain is a simple sum-based argument on depth/z-coordinates rather than a broad proof.
Experimental setup
Datasets: Real3D-AD Anomaly-ShapeNet ModelNet40 ShapeNet-Part
Benchmarks: Real3D-AD Anomaly-ShapeNet ModelNet40 ShapeNet-Part
Metrics: O-AUROC P-AUROC Accuracy IoU FPS
Baselines: BTF M3DM PatchCore CPMF RegAD IMRNet
Builds on: PointMAE RegAD PatchCore FPFH PointNet++
Closest papers by meaning
- Align3D-AD: Cross-Modal Feature Alignment and Dual-Prompt Learning for Zero-shot 3D Anomaly Detection (NeurIPS 2026 · similarity 0.85)
- Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection (CVPR 2025 · similarity 0.85)
- Rethinking Reverse Distillation for Multi-Modal Anomaly Detection (AAAI 2024 · similarity 0.85)
- Towards Scalable 3D Anomaly Detection and Localization: A Benchmark via 3D Anomaly Synthesis and A Self-Supervised Learning Network (CVPR 2024 · similarity 0.84)
- Beyond Single-Modal Boundary: Cross-Modal Anomaly Detection through Visual Prototype and Harmonization (CVPR 2025 · similarity 0.84)
- Revisiting Multimodal Fusion for 3D Anomaly Detection from an Architectural Perspective (AAAI 2025 · similarity 0.84)
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.