See a capture end to end
Two capture sets ship with this site, one from each release. Both are complete — every modality, every ground-truth companion file, the calibration and the point clouds — so you can see exactly what a frame contains before downloading anything.
The 1080P export is a single high-resolution frame with a global instance legend. The HD720 release trades resolution for coverage, adding semantic classes, detection labels and three elevation rings.
The two cases
1080P example
/ example_25
1920×1080
example_25 / left / 015_horizontal_15.png- Camera
- Stereolabs ZED Mini (simulated, HD1080)
- Frames here
- 1 across 0°
- Modalities
- 11
- Sample size
- 30 files, 136.8 MB
720P example
/ example_720
1280×720
example_720 / left / 002_horizontal_02.png- Camera
- Stereolabs ZED Mini (simulated, HD720)
- Frames here
- 3 across 0°, +45°, −45°
- Modalities
- 15
- Sample size
- 88 files, 66.0 MB
Side by side
| 1080P example | 720P example | |
|---|---|---|
| Resolution | 1920×1080 | 1280×720 |
| Rendered at | 1920×1080 (native) | 2560×1440, downsampled ×2 |
| Frames in this sample | 1 | 3 |
| Elevation rings | 0° | 0°, +45°, −45° |
| Orbit per tree | 356° sweep, 48 views total | 356° sweep, 16 views/ring, 48 views total |
| Camera distance | 4.00 m, fixed | 2–4 m to the trunk surface, cycling 2, 2.5, 3, 3.5, 4 m |
| Modalities present | 11 | 15 |
| Instance segmentation | Global legend, 5124 instances with fixed colours | Per-frame ids; the target tree is always id 1 |
| Semantic classes | — | 11 classes, 10 present here |
| Detection labels | — | YOLO polygons + boxes, class tree |
| Point clouds | 2 per-frame | 4 merged over the orbit |
| Sample on this site | 30 files, 136.8 MB | 88 files, 66.0 MB |
Every figure above is read from the capture set on disk, so a dash means that release genuinely does not carry the field — not that it is missing from this page.
Which one to start with
The 1080P example
It has the most detail per frame, and its depth, disparity and normals come as float EXR and NPY rather than quantised 16-bit maps. Its instance legend also keeps the same colours across the whole scene, which matters if your method follows instances from one frame to the next.
The 720P example
It ships YOLO polygons and boxes, so detection and segmentation training needs no labelling of your own. Every pixel also carries a semantic class — ground, fern, shrub, deadwood and the rest — where the 1080P export separates tree instances only. The frames are smaller, and the orbit is shot from above and below as well as level.
Get the full release
It repeats every modality shown in these walkthroughs across the complete orbit for each published tree.