Looking 3D: Anomaly Detection with 2D-3D Alignment

Looking 3D: Anomaly Detection with 2D-3D Alignment

Ankan Bhunia, Changjian Li, Hakan Bilen CVPR 2024

Dataset

The BrokenChairs180K dataset is available for download from here.
The dataset contains around 180K rendered images with 100K classified as anomaly and 80K normal.
Each rendered query image is associated with a normal shape reference.
Different types of abnormalities include: missing parts, broken parts, swapped components, mis-alignments.
The query pose is unknown.
Testing is performed on previously unseen instances.

filename and download link	folder structure	size (after extracting)	comments
images.zip	BrokenChairs/images/	21 GB	[1] (see below)
annotations.zip	BrokenChairs/annotations/	2 GB	[2] (see below)
shapes.zip	BrokenChairs/shapes/	14 GB	[3] (see below)
split.json	BrokenChairs/split.json	134 KB	[4] (see below)

Note:

[1]BrokenChairs/images/: The filenames for the images have a specifc structure. For example in the file with name render_183_1944_2.5_300_30_3_normal.png, 183 is the shape_id, 1944 is the texture_id, 2.5_300_30_3 contains info on camera paramters (in the format of <radius>_<azim>_<elev>_<light-index>).

[2]BrokenChairs/annotations/:<info_*>: It contains 2d_bbox, IoU, camera_parameters and texture_id. <mask_new_*>: binary mask of the object part with the anomaly. <mask_old_*>: binary mask of the object part without the anomaly (normal). <mask_new_*>: segmentation mask of the chair with the anomaly. <mask_old_*>: segmentation mask of the chair without the anomaly (normal).

Annotations are available for anomaly images only. For some anomaly types like missing component, <mask_old_*> is not available.

[3]BrokenChairs/shapes/: <mv_images/*.png>: grayscale multi-view image, <mv_images/*.json>: json file containing intristic and extrinsic parameters of the rendered image, <mv_images/*.npy>: npy file containing 2D-3D correspondence points. <model_id.txt>: corresponding ShapeNet id.

Please refer to utils/render_multiview.py which can be used to obtain the above <png/json/npy> files from any given obj/stl/glb mesh shape.

[4]BrokenChairs/split.json: train/test/val split. Each set has mutually exclusive shape instances.

Distribution of anomaly types within our dataset, categorized by salient chair parts, is shown below.

Citation

If you use the results and code for your research, please cite our paper:

@article{bhunia2024look3d,
  title={Looking 3D: Anomaly Detection with 2D-3D Alignment},
  author={Bhunia, Ankan Kumar and Li, Changjian and Bilen, Hakan},
  journal={CVPR},
  year={2024}
}