Event6DBlender (Blender-Rendered Training Data — easy subset)

Synthetic Blender renders + simulated event streams used to train the depth-extrapolation network behind Event6D (CVPR 2026).

This repository hosts the easy subset that was actually consumed by the released training run (the dataloader hardcodes categories=['easy']). A separate companion repo mickeykang/Event6DBlenderMedium hosts the medium extension — extra data not used by the released checkpoint.

Layout

Event6DBlender/
├── train.txt                                # full split list (714 easy + 1354 medium)
├── test.txt                                 # 590 sequences (204 easy + 386 medium)
├── gso/<obj_id>/...                         # 1035 Google Scanned Objects meshes (CC-BY 4.0)
├── EvBlenderProc/25-07-30_easy_9/train_pbr/<seq>/
│   ├── rgb/<frame>.png                      # 120 frames per sequence, 480×640 RGB
│   ├── depth/<frame>.png                    # 16-bit metric depth (scale in scene_camera.json)
│   ├── mask/<frame>_<obj>.png
│   ├── mask_visib/<frame>_<obj>.png         # visible-object masks (used by dataloader)
│   ├── scene_camera.json                    # per-frame intrinsics K
│   ├── scene_gt.json                        # per-frame (R, t) for every object
│   ├── scene_gt_coco.json, scene_gt_info.json
└── EvBlenderProcEv/25-07-30_easy_9/<seq>/
    └── 0001.npz, 0002.npz, ...              # raw events per inter-frame interval
                                             #   .npz['data'] = struct(x, y, t, p)

Per-sequence: ≈120 RGB frames + 120 depth + ≈30 event npz files. 714 sequences total.

Splits

• train.txt: 2068 sequences (714 easy + 1354 medium)
• test.txt: 590 sequences (204 easy + 386 medium)

Released checkpoint uses easy only. For medium, see mickeykang/Event6DBlenderMedium.

Download

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

Disk-space note

The training pipeline materializes a voxel-grid cache next to the events on first run (EvBlenderProcEv_cache/, ≈90 GB across the full split). The cache is deterministic and disposable — delete it any time to reclaim space.

Attribution

• GSO meshes: Google Scanned Objects by Google Research — released under CC-BY 4.0.
• Renders produced with BlenderProc; events simulated with ESIM (UZH-RPG).

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}
}