flw-tu-dortmund/UnrealEvent

⚠️ Work in progress — everything here is still being tested. These datasets, the generation pipeline and the label formats are under active development and not yet validated for production use. Expect labels, folder structure and conventions to change without notice, and verify anything before you rely on it. Feedback is welcome.

FLW Synthetic Event-Camera Warehouse Datasets

Synthetic datasets for event-camera perception in logistics/warehouse scenes. Each dataset pairs high-frame-rate rendered video with a simulated event stream and pixel-exact ground truth (2D boxes, oriented 3D boxes, 6-DoF pose) projected directly from the 3D scene — not produced by a detector.

The dataset viewer is disabled on purpose: these are image sequences plus custom label files (COCO / BOP / YOLO), not a single table the viewer could render.

Datasets

Dataset	Frames	Duration	Resolution	Classes	Label formats
WarehouseBig	10,000	33.33 s	1280×720	human, slc, pallet, pallet_jack	COCO, BOP, YOLO
WarehouseSmall	5,000	16.67 s	1280×720	pallet, barrel, pallet_jack, slc	COCO, BOP
FLW Working Space	1,500	5.00 s	1280×720	robot, human, pallet, pallet_jack, klt	COCO, BOP, YOLO

All sequences are rendered at 300 fps (the event simulator infers motion from the difference between consecutive frames, so a high rate gives clean events).

Examples (RGB frame with projected 2D labels)

WarehouseBig	WarehouseSmall	FLW Working Space

(Boxes are drawn from the COCO labels; a yellow box + [MOVING] marks an object that is actually moving in the scene.)

How the data was made

A fully automated, in-editor pipeline (the Event Camera Toolkit plugin for Unreal Engine 5.6):

1. Render — a cinematic camera (Level Sequence) is rendered to an image sequence with Movie Render Queue at 300 fps, 1280×720, PNG, motion blur off (event simulators want sharp frames).
2. Events — the frames are converted to an event stream with v2e. Settings used for every dataset: --input_frame_rate 300 --dvs_exposure source --pos_thres 0.2 --neg_thres 0.2 --sigma_thres 0.03 --cutoff_hz 300 --leak_rate_hz 0.1 --shot_noise_rate_hz 0.01 --timestamp_resolution 0.001 --disable_slomo. Output: v2e.h5 (raw events) + dvs-video.avi (a 30-fps visualization).
3. Labels — for every frame, each tagged actor is projected from the 3D scene into the camera to produce ground truth: 2D box (amodal + visible), oriented 3D box, 6-DoF pose, per-pixel visibility, occlusion level, and an is_dynamic flag. A depth mask per frame gives the tight visible box and the visible fraction.
4. Verification — overlay videos (boxes on the RGB frames and on the event video) are rendered so the labels can be checked by eye.

Everything is ground truth from the simulator, so labels are exact (subject to the caveats below) — there is no annotation noise.

Folder structure

Each dataset is one top-level folder with the same layout:

<Dataset>/
├── Frames/
│   ├── frame_000000.png …                 # rendered RGB, 300 fps, 1280×720
│   └── labels/
│       ├── coco/instances.json            # COCO (the superset: 2D+3D+6D, see below)
│       ├── bop/                           # BOP 6-DoF pose
│       │   ├── scene_gt.json              #   object poses (cam_R_m2c, cam_t_m2c mm)
│       │   ├── scene_gt_info.json         #   bbox_obj/bbox_visib, px counts, visib_fract
│       │   └── scene_camera.json          #   per-frame cam_K + depth_scale
│       └── yolo/                          # (where present) frame_000000.txt + classes.txt
├── Events/
│   ├── v2e.h5                             # raw event stream (x, y, t, polarity)
│   ├── dvs-video.avi                      # event visualization (30 fps)
│   ├── dvs-video-frame_times.txt          # timestamp per event-video frame
│   └── v2e-args.txt                       # exact v2e command + parameters
└── Verification/
    └── overlay_{rgb,events}_{2d,3d,6d}.mp4 # boxes/pose drawn on frames and events

COCO labels (labels/coco/instances.json)

COCO is the richest format — every annotation carries:

Field	Meaning
bbox	primary 2D box [x, y, w, h], pixels
bbox_amodal	full 2D extent, including occluded parts
bbox_visible	tight box around only the visible pixels
visible_fraction, visible_pixels, amodal_pixels	how much of the object is visible
occlusion_level	0 fully visible … 3 mostly hidden
is_dynamic	the object is actually moving in the scene
instance_id	stable per-object id across frames
bbox3d	oriented 3D box (center, extent, rotation), world space
corners_2d, box_edges_2d	3D box projected to the image
pose_6d	6-DoF object→camera pose, OpenCV/BOP convention (R, t in metres, quaternion)

Each image entry also stores its timestamp (seconds) and cam_K (pinhole intrinsics).

Conventions

• Camera / pose: OpenCV convention (x_cam = R · x_obj + t); pose_6d.t is in metres, BOP cam_t_m2c in millimetres. Intrinsics cam_K = [fx,0,cx, 0,fy,cy, 0,0,1].
• Timing: frames are 1/300 s apart; COCO timestamp and the event stream share the same clock, so events and frames line up.
• Classes: defined per scene (see the table); the same tag can apply to many instances of one class.

Loading

import json, h5py

# labels
coco = json.load(open("WarehouseBig/Frames/labels/coco/instances.json"))
print(len(coco["images"]), "frames,", len(coco["annotations"]), "annotations")

# events
with h5py.File("WarehouseBig/Events/v2e.h5", "r") as f:
    print(list(f.keys()))   # event arrays (x, y, timestamp, polarity)

Frames are Frames/frame_%06d.png; a label's image_id equals the frame number.

Known limitations (being tested)

• Synthetic only — absolute numbers carry a sim-to-real gap; the datasets are best used for relative comparisons (e.g. frames vs events, easy vs hard conditions).
• The animated human's 3D box can be oversized (it currently uses the rest-pose reference bounds). The 2D boxes and the 3D boxes of rigid objects are correct.
• Masks/labels and the event simulation are still being validated; details may change.