Benchmark tasks

Compute disparity from the rectified left/right pair. Foliage is the hard case here: thin branches and repeated leaf texture defeat the window-matching assumptions that work on road scenes. Evaluation is performed in non-occluded regions.

Input

left/ + right/ RGB

Ground truth

left_disparity/*_px.npy (float px), *_px16.png (16-bit)

Metrics

D1-all, EPE, bad-1.0 / bad-3.0

Predict per-pixel metric depth. Because the scene is fully synthetic, the ground truth is exact everywhere — including inside the canopy, where a real depth sensor would return nothing at all.

Input

left/ RGB alone, or the stereo pair

Ground truth

left_depth/*.exr, *_m.npy (metres), *_mm.png (16-bit mm)

Metrics

AbsRel, RMSE, δ<1.25

Estimate world-space surface normals taken directly from the scene geometry rather than differentiated from depth, so they stay sharp across bark ridges and leaf edges.

Input

left/ RGB

Ground truth

left_normal/*.exr, *_world.npy (world space)

Metrics

Mean / median angular error, share within 11.25°

Separate individual tree instances and pick the foreground tree out of the surrounding canopy — instances that interpenetrate and share both colour and texture.

Input

left/ RGB

Ground truth

left_seg/*_id.png plus the *_instances.json legend

Metrics

Mask AP, IoU on the target instance

Reconstruct the geometry of the target tree from the orbit. Per-frame clouds share a coordinate frame, so a merged reference is available for direct comparison.

Input

The full orbit of stereo frames

Ground truth

pointcloud/*.ply per frame, plus merged scene aggregates

Metrics

Chamfer distance, F-score at 1 cm / 5 cm

A submission endpoint with a public leaderboard is currently in private alpha. To request access, please contact the maintainer.

In the meantime the dataset is complete enough to evaluate locally: every metric above can be computed from the files shipped with each capture, with no held-out split withheld from you.