tpoellabauer/YCB-V-RAND

We provide our physically-based renderings of the YCB-V object dataset randomized YCB-V (YCB-V-RAND) with randomized texture, randomized material, and randomized lighting, allowing to benchmark the influence of such variations on the task of 2D object detection and 6DoF object pose estimation. For more details on the dataset, please take a look on our paper referenced below.

Contents

Each <scene_id>.tar.gz is a WebDataset shard with train/, val/, test/ splits (rendered PNGs), plus per-scene transforms_*.json (camera transforms), initial_poses_*_gen.json, and test/scene_camera.json / test/scene_gt.json / test/test_targets.json (BOP-format annotations).

Note: some shards contain a couple of empty initial_poses_*_gen.json placeholder files (0 bytes, no pose data generated for that split), which breaks the Hub's automatic WebDataset viewer. Download and extract to use — the image/annotation data itself is unaffected.

Related datasets

• tpoellabauer/YCB-V-DS — stereo/depth recordings, same object set
• tpoellabauer/YCB-V-LUMA — chroma/luma-key recordings
• tpoellabauer/YCB-LUMA — luminance-keyed localization follow-up

If you find our work useful, please consider citing our work.

@inproceedings{pollabauer2024generalizing,
  title={Generalizing Neural Radiance Fields for Robust 6D Pose Estimation of Unseen Appearances},
  author={P{\"o}llabauer, Thomas and Wirth, Tristan and Weitz, Paul and Knauthe, Volker and Kuijper, Arjan and Fellner, Dieter W},
  booktitle={International Symposium on Visual Computing},
  pages={300--314},
  year={2024},
  organization={Springer}
}