VICO-UoE/OddOneOutAD

⚡ Odd-One-Out: Anomaly Detection by Comparing with Neighbors

Goal: Detecting 'odd-looking' samples in multi-object scene environments.

Odd-One-Out: Anomaly Detection by Comparing with Neighbors
Ankan Bhunia, Changjian Li, Hakan Bilen

The input of the framework is a set of sparse view images of a scene containing multiple objects. We aim to detect 'odd-looking' objects that contain manufacturing errors (e.g., different geometry, texture) or damages (e.g., cracks, fractures).

News

• 06/09/2024 - Codes & models coming soon.
• 06/09/2024 - The dataset is made public via huggingface.

🎯 ToysAD-8K and PartsAD-15K datasets

• The ToysAD-8K and PartsAD-15K dataset are available for download here.
• ToysAD-8K includes real-world objects from multiple categories and PartsAD-15K comprises a diverse set of mechanical object parts.
• Both datasets consist of multiple scene folders, each containing RGB rendered images, masks, and segmentations annotations for each multiview image along with their metadata.
• Different types of abnormalities include: missing parts, broken/fracture/cracks parts, mis-alignments, texture mismatch.
• The datasets are divided into chunks of 5GB. We provide scripts to download both datasets.

Dataset name	Command	Total size	Comments
ToysAD-8K	bash data/download.sh toysAD8K	40 GB	Rendered using Toys4K* shapes (Creative Commons and royalty-free licenses)
PartsAD-15K	bash data/download.sh partsAD15K	94 GB	Rendered using ABC* shapes (MIT827 license)

*This repository does not claim ownership of the shapes in the original dataset. To obtain the original shape data, please refer to their official dataset pages. You can retrieve the shape_ids from .json files in the scene folders.

Examples

🔖 click to preview dataset images

Cite our work!

@article{bhunia2024odd,
  title={Odd-One-Out: Anomaly Detection by Comparing with Neighbors},
  author={Bhunia, Ankan and Li, Changjian and Bilen, Hakan},
  journal={arXiv preprint arXiv:2406.20099},
  year={2024}
}