raulcastar/mare-lunar-baselines

MARE Lunar Rover Baselines

Segmentation and depth models trained on the synthetic MARE Lunar Rover Dataset, recorded in the MARE simulator. They are the baselines of the MARE paper; the training and evaluation code is in MARE-Dataset-Baselines.

All models were trained on maps 000-005, selected on map 006 and scored on map 007, terrain never seen in training. Scores are at the dataset's native resolution.

Folder	Task	Model	Input	Test
depth	monocular depth	Depth Anything V2-Small, fine-tuned	rgb_central	AbsRel 0.027, RMSE 1.63 m, δ<1.25 0.984
object	objects	U-Net, ResNet-34 encoder (ImageNet)	rgb_central	mIoU 0.596
object_deeplabv3plus	objects	DeepLabV3+, ResNet-50 encoder (ImageNet)	rgb_central	mIoU 0.570
object_segformer	objects	SegFormer-B2 (MiT-B2 encoder, ImageNet)	rgb_central	mIoU 0.586
rear_object	objects	U-Net, ResNet-34 encoder (ImageNet)	rear_rgb	mIoU 0.666
rear_shadow	shadows	U-Net, ResNet-34 encoder (ImageNet)	rear_rgb	mIoU 0.849
shadow	shadows	U-Net, ResNet-34 encoder (ImageNet)	rgb_central	mIoU 0.854
shadow_deeplabv3plus	shadows	DeepLabV3+, ResNet-50 encoder (ImageNet)	rgb_central	mIoU 0.804
shadow_segformer	shadows	SegFormer-B2 (MiT-B2 encoder, ImageNet)	rgb_central	mIoU 0.801
spad_object	objects	U-Net, ResNet-34 encoder (ImageNet)	spad	mIoU 0.484
spad_shadow	shadows	U-Net, ResNet-34 encoder (ImageNet)	spad	mIoU 0.709

Each folder holds model.safetensors and config.json (class names, input size, per-class test IoU, training settings, code commit and dataset revision).

Usage

Segmentation models are segmentation-models-pytorch networks; inputs are RGB images resized to input_size (width, height) and normalized with the ImageNet mean and standard deviation.

import json
import segmentation_models_pytorch as smp
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

config = json.load(open(hf_hub_download("raulcastar/mare-lunar-baselines", "object/config.json")))
model = smp.Unet("resnet34", encoder_weights=None, classes=config["classes"])
model.load_state_dict(load_file(hf_hub_download("raulcastar/mare-lunar-baselines", "object/model.safetensors")))
model.eval()

Use smp.DeepLabV3Plus("resnet50", ...) and smp.Segformer("mit_b2", ...) for the other architectures. The depth model is Depth Anything V2-Small fine-tuned for metric depth: load the base model with transformers.AutoModelForDepthEstimation, then depth/model.safetensors; metric depth in metres is 10 / predicted_depth.

With the training repository, mare_baselines.predict runs any of these models on a folder of images:

hf download raulcastar/mare-lunar-baselines --local-dir results/hub
docker compose run --rm train mare_baselines.predict --run object --results /results/hub --images /data

License

The U-Net, DeepLabV3+ and depth weights are released under CC BY 4.0. The SegFormer weights derive from NVIDIA's MiT-B2 ImageNet weights and are subject to the NVIDIA SegFormer license (non-commercial use).

Citation

@misc{castilla2026mare,
  title         = {{MARE}: An Open-Source Lunar Rover Simulator with Shadow Ground Truth for Perception and Autonomy},
  author        = {Castilla-Arquillo, Ra{\'u}l and Olivares-Mendez, Miguel},
  year          = {2026},
  eprint        = {2610.28465},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.28465}
}