Event6DBlenderMedium (Blender-Rendered Training Data — medium extension)

The medium-difficulty extension to mickeykang/Event6DBlender. Not used by the released Event6D checkpoint (the public training run hardcodes categories=['easy']), but included here for users who want to extend training to harder sequences.

Layout

Event6DBlenderMedium/
├── EvBlenderProc/25-07-30_medium_9/train_pbr/<seq>/
│   ├── rgb/, depth/, mask/, mask_visib/, scene_camera.json, scene_gt.json, ...
└── EvBlenderProcEv/25-07-30_medium_9/<seq>/0001.npz, 0002.npz, ...

1354 training sequences + 386 test sequences (medium-difficulty category).

Usage

Download into the same local directory as Event6DBlender so the splits and GSO meshes from that repo cover the medium sequences as well:

huggingface-cli download mickeykang/Event6DBlender       --repo-type dataset --local-dir ./data/Event6DBlender
huggingface-cli download mickeykang/Event6DBlenderMedium --repo-type dataset --local-dir ./data/Event6DBlender

train.txt / test.txt from the primary repo already list these sequences.

To train on easy + medium, modify the dataloader call (blender_reader_e2e.py:89) to process_subdir_fast(categories=['easy', 'medium']).

Attribution

Renders use the same GSO mesh set + BlenderProc pipeline as the primary repo — see mickeykang/Event6DBlender for licensing details.

Citation

@inproceedings{kang2026event6d,
  title     = {Event6D: Event-based Novel Object 6D Pose Tracking},
  author    = {Kang, Jae-Young and
               Cho, Hoonehee and
               Lee, Taeyeop and
               Kang, Minjun and
               Wen, Bowen and
               Kim, Youngho and
               Yoon, Kuk-Jin},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}