vveicao/unrealengine

Unreal Engine Outdoor 10K

10,011 synthetic outdoor videos rendered with Unreal Engine 5.7 in four city maps, with per-frame RGB, metric depth, camera intrinsics and poses, instance masks and per-frame 3D oriented boxes (street furniture, trees and every car, including traffic that drives through the shot). Same format as vveicao/astra-collect-20k and vveicao/infinigen-collect-10k: the native scene-JSON format of SpatialEncoder v2 Small (spatial_encoder_v2_small), plus WebDataset shards that its HF/WDS training loader can stream. Only the source name differs (UnrealEngine, root $UNREALENGINE_ROOT).

	train	val	total
Videos	9,019	992	10,011
Frames	541,140	59,520	600,660
Object tracks			945,318 (55 categories)
WDS shards / samples	421 / 9,019	42 / 992	466.4 GB

One folder per map, named by its video count:

Folder	Map	UE project / level	train	val	videos
downtown_west_1k	Downtown West	DowntownWest /Game/Downtown_West/Maps/Demo_Environment	862	138	1,000
small_city_5k	City Sample Small City	CitySample /Game/Map/Small_City_LVL	4,374	626	5,000
city_park_2k	City Park	CityPark /Game/CityPark/Maps/Showcase	2,011	0	2,011
big_city_2k	City Sample Big City	CitySample /Game/Map/Big_City_LVL	1,772	228	2,000

• 60 frames per video, 1024 × 1024, nominal 10 fps (only the order of frames matters to the model).
• Field of view (square, horizontal = vertical) fixed per video, p10 / median / p90 47.1 / 61.6 / 124°; pinhole, no distortion.
• Start-to-end camera distance per video, p10 / median / p90 0.571 / 1.83 / 14.1 m.
• The split follows the camera sequence a video was cut from (about 10 % val), so neighbouring videos never straddle train and val. city_park_2k has only 7 camera sequences, none of which fell into val, so it is train only.
• Pedestrians are removed from the maps. All released videos passed a pedestrian detector check (Faster R-CNN, every 5th frame, then a visual check); 51 videos with a remaining distant pedestrian or a dropped frame were removed (metadata/selections/dropped_and_topup.json).
• Most frequent categories (tracks): bollard (132,482), tree (116,795), car (114,006), street light (89,373), fence (83,398), parking meter (49,723), litter (46,710), barricade (41,305), sign (31,558), bench (27,460).

Layout

README.md
manifest.json                         WDS shard manifest (schema spatialencoder-wds-native-v1, pinned URLs)
catalog.json                          training catalog with the "UnrealEngine" source ($UNREALENGINE_ROOT paths)
dataset/UnrealEngine.json              index: [["train-json/<map folder>/<video>.json", num_frames, "train"|"val"], ...]
UnrealEngine/train-json/<map folder>/<video>.json   one scene JSON per video (all splits; the index gives the split)
videos/<map folder>/<video>.tar       per-video package, extracts to <video>/...
  <video>/render/rgb_frames/NNNNNN.jpg   RGB
  <video>/render/rgb.mp4                 the same frames as a video
  <video>/depth/depth_mm_uint16/NNNNNN.png   depth, uint16 millimetres, 0 = invalid (sky or farther than 65.535 m)
  <video>/depth/depth_preview.mp4
  <video>/instance/instance_id_png/NNNNNN.png   instance ids (uint16, 0 = unannotated)
  <video>/annotations/                   video_annotation.json, per_frame/*.json, categories, source
  <video>/trajectory.json                camera path of the video
wds/UnrealEngine/unrealengine-<split>-NNNNNN.tar   ~1 GB WebDataset shards (rgb + depth + metadata)
metadata/registry.jsonl               every video: map, source sequence, split, end cameras, FOV, object counts
metadata/selections/                  the released video list, folder of each video, dropped videos and top-ups
metadata/stats.json                   counts and the category histogram
metadata/videos_sha256.json           size and sha256 of every videos/*.tar

All tars extract into one videos/ folder, so the frame paths in the scene JSON need no change. Video ids are ue_<map>_<sequence>_<index>.

Conventions

Identical to the Astra and Infinigen releases. Scene JSON frames[frame_id]:

Field	Meaning
rgb_path, depth_path, instance_path	relative to videos/ (after extracting the tars)
rgb_intrinsic, depth_intrinsic	3 × 3 K in pixels, pixel centres at integer coordinates (cx = cy = 511.5); fx = fy, per video
camera_pose	4 × 4 camera-to-world, metres, OpenCV camera axes (+X right, +Y down, +Z forward), world +Z up (Unreal world with Y flipped)
original_width, original_height	1024, 1024
time_s	frame time at the nominal 10 fps

Scene JSON objects[object_id][frame_id] (only frames where the object is visible):

Field	Meaning
camera_center	box centre in the camera frame, metres
camera_scale	full side lengths, metres, in the order of camera_R's columns
camera_R	box-local → camera rotation; local Y is the box axis closest to world vertical, pointing down
camera_box_2d_rend	2D prompt / 2D target used by the loader: envelope of the projected cuboid, clipped to the image [-0.5, W-0.5] × [-0.5, H-0.5]
camera_box_2d_visible	box of the visible instance mask
camera_box_2d_proj	envelope of the projected cuboid, not clipped
category_name, visible_pixels	query text and visible pixel count

Boxes come from the meshes placed in the level (static props) and from the traffic actors logged at every frame (vehicles), so moving cars have exact per-frame boxes. Instance masks are the pixels whose back-projected depth falls inside a box (props also match the renderer's per-mesh stencil id); where boxes overlap the nearer box wins.

Source content

Rendered from Epic Games' City Sample (Small City, Big City) and the Downtown West and City Park environments from Fab. Check the Unreal Engine EULA and the Fab content licences before use.