RAD is a robot-captured benchmark for anomaly detection under changing viewpoints, uncontrolled illumination, reflective materials, and viewpoint-dependent defect visibility. It contains 4,287 RGB images, 13 semantic object categories represented by 18 physical instances, and four defect types: scratched, missing, stained, and squeezed. Each object instance is captured from 68 predefined robot viewpoints with a Franka Emika Panda and Intel RealSense D415.

News

• 2026-07: Public benchmark adapters, validation tools, and project page released.
• 2024-10: RAD first appeared on arXiv.

Overview

RAD targets pose-agnostic anomaly detection: training uses only defect-free views and a test image arrives at an unknown pose. A method must separate a real defect from normal changes caused by pose, shading, reflection, or partial occlusion. The paper evaluates three families:

• 2D feature methods: CFlow, EfficientAD, FastFlow, PaDiM, PatchCore, Reverse Distillation, STFPM, UFlow, WinCLIP, AdaCLIP, and VCP-CLIP.
• 3D reconstruction methods: SplatPose and PIAD, using COLMAP-estimated camera poses from unmasked images rather than robot ground-truth poses.
• Vision-language models: Qwen2.5-VL and GPT-4o, evaluated with classification, bounding-box localization, and box-to-mask conversion.

The main result is that mature 2D feature methods remain strongest for image-level detection; VCP-CLIP leads pixel-level AUROC, while the 3D methods are competitive at localization but remain sensitive to reflection, symmetry, and sparse views.

Dataset

Hosted downloads:

• RAD with pixel masks
• RAD without masks

The paper reports 13 semantic categories: binder clip, bowl, box, can, charger, cup 1, cup 2, glue bottle, phone case, rubber duck, spoon, spray bottle, and tennis ball. Multiple physical instances of binder clip, cup 2, glue bottle, and phone case are kept in separate directories and aggregated into their semantic category for reporting.

RAD_with_mask/
├── binderclip/
│   ├── train/good/                 # normal training views
│   ├── test/good/                  # normal test views
│   ├── test/{missing,...}/         # anomalous test views
│   ├── ground_truth/{missing,...}/ # pixel masks
│   └── transforms.json             # camera metadata when included
├── binderclip2/
└── ...                             # 18 physical-instance directories

After extracting an archive, validate paths, masks, counts, and optional pose metadata:

rad-benchmark validate /path/to/RAD_with_mask
rad-benchmark validate /path/to/RAD_with_mask --require-poses --output validation.json

The hosted download links are retained from the earlier repository. The code license in this repository does not automatically grant rights to the dataset; consult the dataset distribution and authors for its applicable terms.

Setup

The release utilities require Python 3.10 or newer and do not require a GPU:

git clone https://github.com/kaichen-z/RAD.git
cd RAD
conda env create -f environment.yml
conda activate rad-benchmark

Or use a virtual environment:

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -e .

Heavy baselines need their own CUDA-compatible environments. Fetch only the method you intend to run; pinned revisions and license notes live in methods/manifest.toml.

Quick start

Generate portable metadata for CLIP methods:

rad-benchmark metadata /path/to/RAD_with_mask work/rad-meta.json
python methods/adaclip/rad_dataset.py work/rad-meta.json --summary

Run one paper baseline through the direct Anomalib adapter:

pip install -e '.[anomalib]'
rad-benchmark anomalib /path/to/RAD_with_mask outputs/anomalib \
  --model patchcore --instances binderclip

Convert provider-neutral VLM box responses into pixel masks:

rad-benchmark vlm-masks responses.json /path/to/RAD_with_mask outputs/vlm-masks

See methods/README.md for CLIP and 3D setup, pinned upstream revisions, and license boundaries. No script embeds API keys, downloads model weights implicitly, or assumes a machine-specific path.

Paper benchmark

Mean AUROC values below are transcribed from the current paper tables.

Family	Method	Image AUROC	Pixel AUROC
Feature	CFlow	0.618	0.984
Feature	EfficientAD	0.803	0.940
Feature	FastFlow	0.786	0.912
Feature	PaDiM	0.587	0.972
Feature	PatchCore	0.833	0.978
Feature	Reverse Distillation	0.784	0.984
Feature	STFPM	0.679	0.973
Feature	UFlow	0.608	0.951
Zero-shot	WinCLIP	0.646	0.853
Zero-shot	AdaCLIP	0.627	0.960
Zero-shot	VCP-CLIP	0.663	0.987
3D	SplatPose	0.524	0.977
3D	PIAD	0.634	0.984
VLM	Qwen2.5-VL	0.305	0.538
VLM	GPT-4o	0.517	0.517

Image- and pixel-level AUROC are the primary metrics. Semantic-category results aggregate the corresponding physical instances; avoid treating the 18 directories as 18 paper categories.

What changed from the research repository

This public release preserves the benchmark protocol while replacing machine-specific research code with small, auditable adapters:

• removed full copies of third-party projects, nested submodules, checkpoints, logs, generated results, notebooks, editor settings, and unused AnySplat/PAD/NeRF experiments;
• added a direct Anomalib Folder adapter for the eight feature models and WinCLIP;
• added root-independent AdaCLIP metadata loading and a collision-safe VCP-CLIP converter;
• added non-destructive SplatPose preparation and a portable PIAD config adapter; because the PIAD project-page repository does not currently contain method source, this release does not claim an executable PIAD reproduction;
• added a provider-neutral VLM JSON prompt and deterministic box-to-mask implementation;
• added validation, smoke tests, CI, contribution/security policy, citation metadata, and explicit upstream revisions and licenses.

These are integration changes only; RAD does not relicense or redistribute third-party method implementations or weights.

Citation

@misc{zhou2024rad,
  title         = {RAD: A Realistic Multi-View Benchmark for Pose-Agnostic Anomaly Detection},
  author        = {Zhou, Kaichen and Chang, Xinhai and Kim, Taewhan and Zhang, Jiadong and Cao, Yang and Peng, Chufei and Zhan, Fangneng and Zhao, Hao and Dong, Hao and Ting, Kai Ming and Zhu, Ye},
  year          = {2024},
  eprint        = {2410.00713},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2410.00713},
  url           = {https://arxiv.org/abs/2410.00713}
}

License

RAD-authored code in this repository is released under the MIT License. Dataset, upstream method, model-weight, and service-provider terms are separate; see methods/manifest.toml.

Acknowledgements

The benchmark adapters build on Anomalib, AdaCLIP, VCP-CLIP, SplatPose, PIAD, COLMAP, and their underlying open-source dependencies. We thank their authors and the hardware/software communities that make reproducible robotic vision research possible.