The Tree Vision Stereo Benchmark
A project of the School of Engineering and Computer Science, Victoria University of Wellington and the UC Vision Group, University of Canterbury
VUW · ECS UC Vision
Y. Lin · M. Zhang · B. Xue · R. Green · S. Schofield
Log in Register
Synthetic stereo benchmark · arboreal scenes

Ground truth for trees, not traffic.

A high-resolution, physically based stereo benchmark aimed at individual trees and the forests around them. Where KITTI gave the automotive community dense ground truth for streets, this suite delivers the same rigour for natural geometry — the hard case for every modern 3D-perception pipeline.

Each capture orbits a target tree at multiple elevations (0°, ±45°) with a simulated Stereolabs ZED Mini-class rig, and every frame ships with stereo RGB, metric depth, disparity, world-space normals, instance segmentation and a dense point cloud.

See the worked examples Browse the dataset Download
RGB, depth, segmentation and disparity from one flight through forest_new — see it full length

A flight through the forest

Alongside the orbits of single trees, we flew a camera through the whole forest_new scene on one continuous path — 74 seconds, 2220 frames, the same simulated ZED Mini rig. The video shows four views of that flight at once, all from the same instant and the same camera, so you can watch the ground truth move with the picture.

top left

RGB (left eye) ZED Mini sim

The left camera of the simulated stereo rig — the view a drone flying the route would record.

top right

Depth 0 – 65 m

Metric distance to the first surface, false-coloured near to far, with the scale down the side of the pane.

bottom left

Segmentation: trees live tree-pixel %

Which pixels belong to a tree, counted as the flight goes and shown in the badge.

bottom right

Disparity fx·B/Z, B = 63 mm

The horizontal shift between the two cameras — the same quantity the capture sets ship per frame.

HD720

1280×720
Still from the HD720 fly-through

The HD720 encode is not in this checkout — it is too large to keep in the repository. Rebuild it with tools/build_site_video.py.

Same output size as the HD720 capture set. This flight was rendered at 2560x1440 and downsampled 2x.

1080P

1920×1080
Still from the 1080P fly-through

The 1080P encode is not in this checkout — it is too large to keep in the repository. Rebuild it with tools/build_site_video.py.

Same output size as the 1080P capture set, though the flight was rendered at 2880x1620 and downsampled 1.5x, while that capture set is native 1920x1080.

Both files are the same flight, re-encoded for the web from masters of 294 MB and 641 MB. Each channel was also kept on its own at full resolution: left and right RGB, depth, segmentation, disparity, and a depth pass covering only the 2500 labelled trees.

Dataset at a glance

2 Capture sets 2 bundled example
96 Planned captures 4 frame(s) published
118 Files on disk 202.8 MB
15 Modalities
5124 Scene instances
1920×1080 / 1280×720 Resolutions 63.0 mm baseline

What's inside every frame

ModalityDescription
left
Left RGB 		Left-camera RGB renderings of the stereo pair.
right
Right RGB 		Right-camera RGB renderings forming the stereo pair (63 mm baseline).
left_depth
Left Depth 		Per-pixel metric depth. EXR/NPY and a 16-bit millimetre PNG for ground truth, *_vis.png for preview.
right_depth
Right Depth 		Per-pixel metric depth for the right camera.
left_disparity
Left Disparity 		Disparity in pixels (left frame), 16-bit PNG and float NPY.
right_disparity
Right Disparity 		Disparity in pixels (right frame).
left_normal
Left Surface Normals 		World-space surface normals. EXR/NPY for ground-truth, *_vis.png for preview.
right_normal
Right Surface Normals 		World-space surface normals for the right view.
left_seg
Left Instance Segmentation 		Per-instance segmentation: colourised PNG, integer ID PNG, target mask and overlay.
right_seg
Right Instance Segmentation 		Per-instance segmentation for the right view.
left_class
Left Semantic Classes 		Semantic class map: *_class.png holds the integer class per pixel, *_class_color.png is the colourised preview.
right_class
Right Semantic Classes 		Semantic class map for the right view.
left_labels
Left Detection Labels 		YOLO annotations for the target tree: normalised polygons in *_labels.txt, boxes in *_labels_bbox.txt.
right_labels
Right Detection Labels 		YOLO polygons and boxes for the right view.
pointcloud
Point Clouds 		Dense point clouds (.ply), per frame and/or merged across the orbit, in raw and coloured variants.

Want to see them side by side? Compare the two worked examples →

Published capture sets

Citation

The paper is on arXiv at https://arxiv.org/abs/2609.13232. If this dataset is useful in your research, please cite: 

@misc{Lin2026TreeVision,
  author        = {Yida Lin and Mengjie Zhang and Bing Xue and Richard Green and Sam Schofield},
  title         = {The Tree Vision Stereo Benchmark Suite},
  year          = {2026},
  eprint        = {2609.13232},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2609.13232},
  institution   = {Victoria University of Wellington and University of Canterbury},
  note          = {A joint project of VUW ECS and the UC Vision Group}
}

Copyright & licence

All datasets and benchmarks on this site are released for academic, non-commercial use under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 licence. Please attribute the authors when redistributing or building upon this work.

Privacy

The renderings are produced from synthetic 3D foliage assets and contain no personally identifiable information. We record anonymised page views and download clicks to report dataset reach. If anything published here concerns you, please contact us.