README.md

---
license: cc-by-nc-sa-4.0
tags:
- Anomaly
- Detection
- Segmentation
pretty_name: SiM3D

extra_gated_prompt: "You agree not to use the dataset to conduct malicious experiments."
extra_gated_fields:
  Full Name: text
  Affiliation: text
  Country: country
  I agree to use this dataset for non-commercial use only: checkbox
---

# 🔥 SiM3D: Single-instance Multiview Multimodal 3D Anomaly Detection 🔥
The first comprehensive benchmark for true 3D anomaly detection using multiview and multimodal data.
- Project Page: https://alex-costanzino.github.io/SiM3D/;
- Paper: https://huggingface.co/papers/2506.21549 - https://arxiv.org/html/2506.21549v1;
- Accepted at ICCV 2025.

### 🎯 Key Innovations
SiM3D is the first benchmark considering the integration of multiview and multimodal information for comprehensive 3D anomaly detection and segmentation (ADS), where the task is to produce a voxel-based Anomaly Volume. 
Unlike existing 2D benchmarks that evaluate pixel-level anomaly maps, SiM3D evaluates complete 3D anomaly volumes, enabling precise defect localisation in manufacturing environments.

### 🏭 Single-Instance Manufacturing Focus
SiM3D focuses on a scenario of high interest in manufacturing: single-instance anomaly detection, where only one object, either real or synthetic, is available for training. 
This addresses the critical industrial challenge where collecting multiple training samples is costly and time-intensive, especially during line changeovers.

### 🌟 Breakthrough Features

- **Synthetic-to-Real Domain Bridge:** SiM3D stands out as the first ADS benchmark that addresses the challenge of generalising from synthetic training data to real test data. Train on CAD models, test on real objects – a game-changer for cost-effective manufacturing QC;
- **Industrial-Grade Data Quality:** The dataset features multiview high-resolution images (12 Mpx) and point clouds (7M points) for 333 instances of eight types of objects, alongside a CAD model for each type. Captured with top-tier ZEISS Atos Q industrial sensors for unmatched precision;
- **Comprehensive 3D Ground Truth:** We also provide manually annotated 3D segmentation GTs for anomalous test samples. Expert-validated voxel-level annotations enable precise 3D defect evaluation.

### 📊 Dataset Statistics
- **333 object instances** across **8 manufacturing object types**;
- **12-36 views per instance** with poses from concentric hemispheres;  
- **High-resolution data**: 12 Mpx grayscale images + 5-7M point meshes;
- **Two evaluation setups**: real2real and synth2real training scenarios;
- **Manually crafted defects**: dents, scratches, contaminations, and paint modifications.

### 🔬 Why SiM3D Matters
Current 2D anomaly detection methods fall short for manufacturing applications requiring automatic intervention. 
Many industrial applications require precise localisation of defects in the 3D space to allow automatic intervention, reducing waste and optimising production. 
SiM3D enables the development of methods that can pinpoint defects in true 3D space, not just image pixels.
Perfect for researchers developing next-generation industrial quality control systems that can generalise from minimal training data while providing actionable 3D defect localisation.

---

### 📋 Instructions
1. Download the dataset:
```
hf download arcanoXIII/SiM3D --repo-type dataset --local-dir ./SiM3D
```
- The download takes a while since the entire compressed dataset is approximately 400 GB;
- Make sure to be logged in using `hf auth login`.
2. Uncompress the class folders:
```
cd SiM3D
bash uncompress.sh
````
- The script uncompresses and removes the folders right after;
- The uncompression takes a while and consumes a lot of space since the entire dataset is approximately 800 GB.
3. Merge split class folders:
```
bash merge.sh
```

At this point, the dataset is ready to be used with the provided dataloader script.

--- 

### 🩻 Depth Maps Extraction
If you wish, you can render depth maps to be used in place of point clouds for your experiments:

```
bash render_depth.sh
```

Keep in mind that the provided dataloader works with depth maps. An alternative class will be provided to work with point clouds.

--- 

### 📠 Dataloader and Evaluation Script [TBA]


---

### 🗄️ Citation
If you find this research useful, please 🥺 cite us with:
```
@inproceedings{costanzino2025sim3d,
  author    = {Costanzino, Alex and Zama Ramirez, Pierluigi and Lella, Luigi and Ragaglia, Matteo and Oliva, Alessandro and Lisanti, Giuseppe and Di Stefano, Luigi},
  title     = {SiM3D: Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark},
  booktitle = {International Conference on Computer Vision (ICCV)},
  year      = {2025},
}
```