Left RGB left
The reference view — a physically based rendering of the target tree, shot with the simulated ZED Mini left sensor.
- 002_horizontal_02.png — preview
Worked example
720P example
— example_720, tree390
These are 3 frames from the benchmark, shown modality by modality. Each frame was rendered from one instant of the same virtual stereo rig, so its depth, disparity, normals and segmentation are all pixel-aligned with its RGB.
The full release repeats this for 48 viewpoints per tree across a 356.0° orbit; 3 frames and 88 files (66.0 MB) are published here as a sample.
1280×720
Resolution
rendered at 2560×1440
63.0
Baseline (mm)
90.0°
Horizontal FOV
0.1–15.0
Depth range (m)
11
Semantic classes
10 present here
88
Files in this capture
66.0 MB
The stereo pair
Two views of the same scene, 63.0 mm apart. The horizontal shift between them is exactly what the disparity map below encodes.
Modality by modality
Right RGB right
The matching right-sensor frame. Together with the left frame this forms the rectified stereo pair at a 63 mm baseline.
- 002_horizontal_02.png — preview
Depth left_depth
Per-pixel metric depth in metres. The preview is a colourised visualisation; the companion files carry the raw values.
- 002_horizontal_02_vis.png — preview
- 002_horizontal_02_mm.png — ground truth
Disparity left_disparity
Stereo disparity in pixels, supplied as a 16-bit PNG alongside this preview.
- 002_horizontal_02_vis.png — preview
- 002_horizontal_02_px16.png — ground truth
Surface normals left_normal
World-space surface normals encoded as RGB. Useful for shape-from-shading and geometry-consistency losses.
- 002_horizontal_02_vis.png — preview
Instance segmentation left_seg
Per-instance masks that separate the target tree from the surrounding canopy, with an integer ID map beside them.
- 002_horizontal_02_color.png — preview
- 002_horizontal_02_id.png — ground truth
- 002_horizontal_02_mask.png — ground truth
- 002_horizontal_02_overlay.png — ground truth
Semantic classes left_class
Every pixel labelled with one of eleven scene classes — tree, sapling, shrub, fern, groundcover, leaf litter, rock, ground, deadwood and the rest — as an integer map plus this preview.
- 002_horizontal_02_class_color.png — preview
- 002_horizontal_02_class.png — ground truth
Detection labels left_labels
YOLO annotations for the target tree: one normalised polygon per frame in *_labels.txt and the matching axis-aligned box in *_labels_bbox.txt, under the single class 'tree'.
- 002_horizontal_02.txt — ground truth
- 002_horizontal_02_bbox.txt — ground truth
- 026_up45_10.txt — ground truth
- 026_up45_10_bbox.txt — ground truth
- 035_down45_03.txt — ground truth
- 035_down45_03_bbox.txt — ground truth
Point clouds
Each view is also back-projected into a dense point cloud in PLY format — the left and right clouds share a coordinate frame, so they can be merged directly or used to validate a reconstruction.
Calibration & scene metadata
Camera
- Model
- Stereolabs ZED Mini (simulated, HD720)
- Resolution
- 1280 × 720 px
- Rendered at
- 2560 × 1440 px, downsampled ×2
- Baseline
- 63.0 mm
- Focal length
- 2.1 mm
- Sensor width
- 4.2 mm
- Field of view
- 90.0°
- Camera distance
-
2.0–4.0 m
to the trunk surface
cycling 2.0, 2.5, 3.0, 3.5, 4.0 m - Orbit sweep
- 356.0°
- Views per ring
- 16
- Elevation rings
- 0°, 45°, -45°
Intrinsics
- fx
- 640.0
- fy
- 640.0
- cx
- 640.0
- cy
- 360.0
- Depth range
- 0.1 – 15.0 m
- Instance ID
- 1 — allocated per frame; the target tree is always 1
- Frames kept
- 48 of 96 rendered
- Frames with labels
- 37
Disparity d relates to depth Z by
Z = fx · B / d, with
fx = 640.0 px and
B = 63.0 mm.
Frames in this sample
The published files are named by their index in the release
(0_left.png), so what you see here is what you will find in
the download. The orbit position each one came from is below.
| ID | Frame | Ring | Elevation | Capture index |
|---|---|---|---|---|
002_horizontal_02 |
horizontal | 0° | 2 | |
026_up45_10 |
up45 | 45° | 26 | |
035_down45_03 |
down45 | -45° | 35 |
Semantic classes
Alongside the per-instance masks, every pixel in this release carries a
semantic class. The *_class_class.png map stores the integer
index below; *_class_class_color.png is only a preview.
- 0 · background (absent)
- 1 · tree
- 10 · other
- 2 · sapling
- 3 · shrub
- 4 · fern
- 5 · groundcover
- 6 · leaf_litter
- 7 · rock
- 8 · ground
- 9 · deadwood
The detection labels are separate and much simpler: a single class,
tree (class 0),
with one normalised polygon per frame for the target tree in
*_labels.txt and the matching axis-aligned box in
*_labels_bbox.txt.
20 tree instances appear in this
capture's segmentation, of which
tree390 is the orbit target.
The rest of the release
Every modality above repeats across the complete orbit for each published tree.