BennyLin01's picture
Add dataset card; point viewer at index.csv so tar shards are not parsed as WebDataset
9a54d36 verified
|
Raw History Blame Contribute Delete
6.01 kB
metadata
pretty_name: Real World Virtual Tree (forest_720)
task_categories:
  - depth-estimation
  - image-segmentation
  - object-detection
tags:
  - stereo
  - synthetic
  - unreal-engine
  - forestry
  - tree
  - point-cloud
size_categories:
  - 1K<n<10K
configs:
  - config_name: frame_index
    data_files: index.csv
    default: true

Real World Virtual Tree (forest_720)

Synthetic stereo captures of individual trees in a procedural forest, rendered in Unreal Engine 5 at 1280x720 per eye. Each frame comes with depth, disparity, surface normals, semantic and instance segmentation, class maps, YOLO labels and a merged point cloud.

287 trees, 8,610 stereo frames, 259,809 files, 153.5 GiB.

The dataset viewer above shows only index.csv, the frame index. The imagery is stored in 45 tar shards (see Layout); they are plain directory archives, not WebDataset, so the viewer deliberately ignores them. Use the download instructions below to get the images.

Capture setup

Values below come from dataset_params.json.

Parameter Value
Resolution per eye 1280 x 720
Render resolution 2560 x 1440 (supersample 2.0)
Stereo baseline 63.0 mm
Focal length 2.1 mm
Sensor width 4.2 mm
Depth range 0.1 - 15.0 m
Camera distance from trunk surface 2.0 - 4.0 m
Views per ring 16
Elevations 0, +45, -45 degrees
Orbit sweep 356 degrees
Views per tree up to 48 (16 x 3)
Source map /Game/Levels/PCG/Breakdown_Levels/forest_new

Tree meshes come from Quixel Megascans 3D Plants.

Semantic classes

0 background, 1 tree, 2 sapling, 3 shrub, 4 fern, 5 groundcover, 6 leaf_litter, 7 rock, 8 ground, 9 deadwood, 10 other.

The YOLO split uses a single class, 0: tree.

Download and extract

Every tar member keeps its path relative to the dataset root, so extracting all shards into one directory reproduces the original layout.

hf download BennyLin01/real_world_virtual_tree --repo-type=dataset --local-dir ./forest_720
cd forest_720 && for f in *.tar; do tar -xf "$f"; done

On Windows PowerShell:

Get-ChildItem ./forest_720/*.tar | ForEach-Object { tar -xf $_.FullName -C ./forest_720 }

To fetch one channel only, pass a pattern:

hf download BennyLin01/real_world_virtual_tree --repo-type=dataset --include "left_seg_*.tar" --local-dir ./forest_720

Layout

After extraction. File names follow <id>_<folder><suffix>, where id is the frame id from index.csv.

Directory Contents Files
left/, right/ stereo RGB, <id>_left.png 8,610 each
left_depth/, right_depth/ _depth_mm.png (16-bit mm), _depth_vis.png, .exr 17,267 each
left_disparity/, right_disparity/ _disparity_px16.png (16-bit), _disparity_vis.png 17,196 each
left_normal/, right_normal/ _normal_vis.png, .exr 8,669 each
left_seg/, right_seg/ _seg_color.png, _seg_id.png, _seg_mask.png, _seg_overlay.png 34,392 / 34,388
left_class/, right_class/ _class.png (class ids), _class_color.png 17,196 / 17,194
left_labels/, right_labels/ _labels.txt, _labels_bbox.txt 17,196 / 17,194
left_cls_src/, right_cls_src/ class source maps 710 each
left_seg_src/, right_seg_src/ segmentation source maps 495 each
pointcloud_merged/ merged point clouds, .ply 1,140
yolo/ data.yaml, images/{train,val}, labels/{train,val} 14,913
capture_tree*.log per-tree capture logs 298

YOLO splits: 5,955 train and 1,501 val images, with matching label files.

Unpacked files at the repo root

File Contents
index.csv frame index: id, tree, tree_index, frame (8,610 rows)
meta.json per-tree geometry: mesh path, centre, extent, base height, tree height
dataset_params.json full capture configuration
progress.json capture progress state
tar_manifest.json per-shard file count and byte size

Shards

45 uncompressed tar shards, each at most about 5 GiB. Counts and sizes are from tar_manifest.json.

Shard Files Size
left_000.tar - left_003.tar 8,610 12.5 GiB
right_000.tar - right_003.tar 8,610 12.5 GiB
left_depth_000.tar - left_depth_001.tar 17,267 8.2 GiB
right_depth_000.tar - right_depth_001.tar 17,267 8.2 GiB
left_disparity_000.tar - left_disparity_001.tar 17,196 5.3 GiB
right_disparity_000.tar - right_disparity_001.tar 17,196 5.3 GiB
left_normal_000.tar - left_normal_003.tar 8,669 18.5 GiB
right_normal_000.tar - right_normal_003.tar 8,669 18.5 GiB
left_seg_000.tar - left_seg_003.tar 34,392 16.4 GiB
right_seg_000.tar - right_seg_003.tar 34,388 16.4 GiB
left_class.tar / right_class.tar 17,196 / 17,194 1.8 GiB each
left_labels.tar / right_labels.tar 17,196 / 17,194 0.14 GiB each
left_cls_src.tar / right_cls_src.tar 710 each 0.15 GiB each
left_seg_src.tar / right_seg_src.tar 495 each 0.03 GiB each
pointcloud_merged.tar 1,140 4.5 GiB
yolo_000.tar - yolo_002.tar 14,913 14.6 GiB
capture_logs.tar 298 0.05 GiB

Notes

  • Tars are uncompressed on purpose: PNG and EXR are already compressed.
  • yolo/data.yaml carries an absolute path: from the capture machine (E:/UE5/forest_720/yolo). Point it at your own extraction directory before training.
  • The data is packed as archives rather than individual files because a Hugging Face repository directory is limited to 10,000 files, and several channels hold far more than that.

License

Not specified yet. Note that the renders derive from Quixel Megascans assets, whose licence terms govern redistribution of derived imagery; set a licence that is consistent with them before treating this dataset as freely reusable.