Event6D Dataset (Real-World Capture)

Real-world hand-held capture used as the primary test set in Event6D (CVPR 2026).

Hand-held capture of 16 daily YCB-style objects from random viewpoints and motions.

Splits

Split	Sequences	Objects	List
test	14 (one per object)	14	test.txt
train	37	16 (adds clamp, pudding)	train.txt

The paper evaluates on the test split only; models are trained on the synthetic EventBlender6D data and never see the real train split. Train sequences use the same capture setup and file layout as test.

Known issue: in banana_1101/0000, the GT rotation flips briefly around frames 252–255 (the banana is nearly symmetric about its long axis and the motion-capture rigid body was ambiguous for a few samples). Skip those frames if you train on rotation labels.

Hardware

Sensor	Model	Resolution	Rate
RGB-D	Intel RealSense D435i	1280 × 720	30 FPS
Event	Prophesee IMX636	1280 × 720	≥ 5000 FPS

Calibration (intrinsics + RGB↔event extrinsic) is shipped as 0001-camchain.yaml in the Kalibr camchain format.

Layout

Event6D/
├── 0001-camchain.yaml        # Kalibr-format intrinsic + extrinsic calibration
├── test.txt, train.txt       # split lists (sequence frame ranges)
├── simple_mesh/              # 21 YCB-style object meshes (textured.obj + material)
└── <object>_<date>/<run>/    # e.g. banana_1101/0003
    ├── parsed_events/        # raw events as packed .npz per frame (x,y,t,p)
    ├── rgb/                  # color frames at 30 fps
    ├── depth_aligned_to_color/
    ├── depth_aligned_to_event/
    ├── pose/                 # GT object poses (4, 4, 4): 4 sub-timesteps per file
    ├── mask/
    ├── obj.txt               # object id string (e.g. 011_wine_glass)
    ├── startend.txt
    └── aligned_depth_pose.csv

Ground-truth pose format

Although RGB is captured at 30 FPS, the event camera produces a continuous stream that we sample at 120 Hz. To give pose labels at the finer rate, each pose/XXXXXX.npy stores 4 poses instead of one — these are 4 evenly-spaced sub-timesteps that span the interval between RGB frame XXXXXX and the next RGB frame XXXXXX+1.

file shape: (4, 4, 4)
              │  └──┴── 4×4 SE(3) rigid-body transform (object → camera)
              └──── sub-timestep index 0..3 within one 30 FPS frame interval

• pose[0] is co-temporal with rgb/XXXXXX.jpg (the start of the interval).
• pose[1], pose[2], pose[3] are 1/4, 2/4, 3/4 of the way to rgb/XXXXXX+1.jpg.

Evaluators can therefore choose either:

• 30 FPS — use only pose[0] per frame, comparing against the tracker's prediction at the RGB instant; or
• 120 FPS — use all 4 sub-timesteps, comparing against pose predictions emitted at each event sub-step.

Event file timing

Events are pre-packed into parsed_events/XXXXXX.npz. Each file holds the events that occurred during the inter-frame interval ending at RGB frame XXXXXX — i.e. the events between rgb/(XXXXXX-1).jpg and rgb/XXXXXX.jpg (a window of ≈ 1/30 s).

time →
    ┌────────────────── 1 / 30 s ──────────────────┐
    rgb/(i−1).jpg                          rgb/i.jpg
    │                                              │
    │←── parsed_events/i.npz  (events in interval) │
    │                                              │
    ├──── pose/(i−1).npy[0..3] (4 sub-timesteps) ──┤

So for any frame i:

• rgb/i.jpg — snapshot at the end of the interval
• parsed_events/i.npz — events leading up to that snapshot
• pose/(i−1).npy[0..3] — 4 GT poses across the same interval, aligned with the events

Each .npz stores a structured array under key data with fields (x, y, t, p) (event pixel coords, timestamp in seconds, polarity ±1).

Download

huggingface-cli download mickeykang/Event6D --repo-type dataset \
    --local-dir ./data/Event6D

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}
}