lothanspace/stride-moon-seg-v1

STRIDE moon segmentation v1

Synthetic stereo frames of the lunar surface (OmniLRS 2.5 / Isaac Sim 5.0, path traced, NASA LOLA Site20 DEM with 2.5 cm streamed high-res terrain), rendered for near-field rock / hazard segmentation from a small rover.

Preview

Left-camera RGB across sun elevations (1-56 deg) and camera heights (0.35-1.0 m):

Everything that ships with one frame (left/right RGB, rock labels, semantic map, traversability, hazard bits, depth, normals):

Rock instances are labelled small/large by measured height above the local ground, split at the wheel-clearance threshold (orange = rock_small, red = rock_large):

The frames used are listed in preview/preview_frames.json.

Clean set

manifest.json lists every frame that passed all checks (4923 of 10925 rendered, 348 terrain locations). splits/<size>_{train,val}.txt are nested by terrain seed (S c M); train/val split by terrain seed hash (15 % val). Available now: S, M. The L size (10k frames) will be released later.

size	train	val
S	1715	289
M	4149	774

rejected.json maps every other rendered frame to its reason; those files are still in the shard folders.

reason	frames
mesh_issue	2969
rock_issue	1472
far_flat	742
dark	302
far_horizon	254
dark_near	227
label_issue	27
overexposed	9

• mesh_issue: rendered terrain more than 30 cm off the DEM within 15 m (stock OmniLRS streaming; rocks appear to float with the ground)
• rock_issue: a labelled rock (or unlabelled pebble) floating more than 6 cm above the terrain within 15 m
• dark_near: less than 50 % of the near field lit
• label_issue: more than 30 % of the pixels carry a rock label and under 10 % ground (no usable ground in view: either the terrain mesh carried the rock label on a few early terrains, or a boulder fills the frame)
• overexposed: terrain mean grey >= 190.0 with too few grey levels to recover (frames above the threshold that WERE recoverable are tone-curved in place: 1094 frames, parameters in meta/<id>.json exposure_fix, originals in images_raw/)
• far_flat / far_horizon: far coarse terrain visible (shards below 1335 rendered a 5 m far mesh; later shards render space beyond ~200 m)

Per-frame files (shards/<shard>/..., <id> = s<shard>_t<terrain>_f<frame>)

• images/<id>_{L,R}.png RGB 1640x1232 (clean render)
• images_sensor/<id>_{L,R}.png the same frame with the sensor model applied (see Sensor realism set)
• labels/<id>_{L,R}_sem.png uint8: 0 space, 1 ground, 2 rock_small, 3 rock_large (split at wheel clearance, 0.05 m placeholder; the rock height is measured from depth + camera pose per blob and stored in meta/<id>.json cams.*.rocks_measured, so the split can be recomputed for another clearance without the depth files)
• labels/<id>_{L,R}_trav.png uint8: 0 free, 1 caution, 2 hazard
• masks/<id>_{L,R}_bits.png bitfield: 1 slope_caution 2 slope_hazard 4 crater 8 rock_small 16 rock_large 32 unlabelled
• depth/<id>_{L,R}_mm.png uint16 millimetres (saturates at 65.535 m; treat >= 65.5 m as far/ignore)
• normals/<id>_{L,R}.png camera-space normals
• meta/<id>.json sun (elevation/azimuth/intensity), rig pose and height, per-camera rocks_measured (px, depth, height above ground, class). The generator's cams.*.rocks height/footprint values are unreliable and kept only for provenance.
• quality.json, raw_manifest.json (intrinsics, terrain seeds), terrains/ (DEM crops, crater tables)

Sensor realism set

The renders are noise-free. shards/<shard>/images_sensor/<id>_{L,R}.png is the same frame passed through a physically based sensor model in post (labels untouched): exposure-time and auto-exposure variation, vignetting, defocus and motion blur, photon shot noise, temperature-dependent dark current, PRNU/DSNU fixed pattern, read noise, 12-bit quantisation. One draw per frame (both cameras share exposure/temperature/blur), seed = crc32(frame id), parameters logged in shards/<shard>/sensor_params.json.

For training, prefer re-randomising per epoch instead of the fixed set: tools/sensor_model.py is numpy/OpenCV only (SensorModel.from_json("tools/sensor_default.json")(img, seed=...), ~0.4 s per 1640x1232 frame, usable in an albumentations Lambda). The shipped calibration is a generic 12-bit CMOS placeholder; fit the real camera with tools/sensor_calibrate.py from dark frames (lens capped, several exposure times at 2+ temperatures) and flat fields (uniform target, exposures from ~5 % to 90 % of saturation), which yields black level, read noise, conversion gain, full well, dark current and its temperature doubling, PRNU/DSNU and the vignetting polynomial.

Background and related work

This dataset grew out of our work with the artificial lunar landscape dataset from Keio University's Space Robotics Group (Ishigami Lab), which we used to train terrain segmentation models for lunar rover navigation. That work shaped our understanding of the problem and informed the class structure used here.

No images, masks, or code from that dataset are included in or derived from this release. This dataset is generated independently using OmniLRS. The two use different rendering pipelines and are intended to be complementary rather than interchangeable: theirs renders procedural landscapes in Terragen with single-camera sky / small-rock / large-rock masks, while this one path-traces real LOLA elevation data in Isaac Sim and adds stereo pairs, metric depth, surface normals, traversability and crater/slope masks.

If you use this dataset, please also consider citing the original:

Pessia, R., Ishigami, G. & Jodelet, Q. Artificial Lunar Landscape Dataset. Kaggle, 2019. https://www.kaggle.com/datasets/romainpessia/artificial-lunar-rocky-landscape-dataset

License and attribution

Released under CC BY 4.0: any use, including commercial products and services, is allowed provided you credit the source. Please cite as: Lothan Space, "STRIDE Moon Segmentation v1", 2026, https://huggingface.co/datasets/lothanspace/stride-moon-seg-v1. Third-party inputs: NASA LOLA elevation data (public domain); terrain texture and rock meshes from the OmniLRS asset set (see that repository for their licences); rendered with OmniLRS 2.5 on Isaac Sim 5.0.

Camera

fx=fy=? px, 1640x1232, hfov 90.0 deg, baseline 0.12 m (placeholder values; replace with the flight camera's). Camera height above ground 0.35-1.0 m, pitch 5-22 deg, sun elevation 2-45 deg weighted to low angles.