Hard3DBench

Sample lists, failure-category labels and human correspondence clicks of Hard3DBench, a diagnostic benchmark with a detailed failure taxonomy under which current feed-forward reconstruction models degrade substantially relative to standard evaluation samples. Introduced in Click3R.

Split	Source datasets	2-view pairs	5-view tuples
In-Distribution	BlendedMVS, CO3D, Habitat, WildRGB-D	404	400
Out-of-Distribution	7Scenes, DTU, ETH3D, NRGBD, ScanNet	300	300

File	Contents
two_view_indist.json, two_view_ood.json	2-view pairs
five_view_indist.json, five_view_ood.json	5-view tuples
categories.json	Failure categories of each 2-view pair (large viewpoint, moderate viewpoint, low overlap, large baseline, extreme depth range, other)
human_clicks.json	Human correspondence clicks on 100 Out-of-Distribution pairs
ood_checksums.json	Checksums of the prepared Out-of-Distribution data
reference.json	Results reported in the paper, unrounded

The metadata of each sample list describes its fields and how the samples were selected. The code repository prepares the source datasets and runs the evaluation.

License

CC BY 4.0 covers our sample lists, labels and clicks only. Images, depth and camera data are not redistributed here — they are referenced by dataset, scene and frame, and remain under the original terms of each source dataset.

Citation

@inproceedings{gu2026click3r,
  title={Click3R: Interactive Stereo 3D Reconstruction with Sparse Correspondence Clicks},
  author={Gu, Hao and Yao, Jin and Cheng, Zezhou},
  booktitle={Neural Information Processing Systems (NeurIPS)},
  year={2026}
}