MarsFM improves Martian terrain maps from single images
MarsFM: Shading-Regularized Flow Matching for Martian Relief Estimation
Computer Vision and Pattern Recognition
Summary
Mapping the surface shapes of Mars from images is tricky because images alone don't easily show height details. The authors designed MarsFM, a computer method that guesses local Martian terrain shapes using single red-band images combined with learned mountain and shadow patterns. They check their guesses against known references and find the method closely matches expected terrain shapes even with fewer computation steps. This approach could help produce better 3D maps of Mars without needing multiple images from different angles.
What this means in practice
- •For planetary mapping teams: Generate improved 3D terrain models of Mars from single-orbit HiRISE images using shading and learned terrain priors.
- •For computer vision engineers: Implement image-conditioned flow matching techniques combining geometric and shading constraints to estimate fine terrain relief from limited spectral image bands.
Authors
Marius F. R. Juston
Abstract
We present MarsFM, an image-conditioned latent flow-matching model for local Martian relief estimation from single-band HiRISE RED orthoimagery. The method combines a pretrained generative prior with stereo-derived geometric supervision and a differentiable Lunar--Lambert shading objective. Relief, normal, gradient, curvature, and ordinal terms constrain complementary aspects of terrain structure, while a positive-affine-invariant image comparison constrains rendered appearance. An evaluation comprising 2024 gathered patch records per integration-step count yields mean affine-aligned RMSE between 0.0935 and 0.0957 in normalized signed-log relief space for one to twenty Euler steps. These scores measure agreement with VAE-reconstructed references on positive-reference support. Their narrow range supports low-step inference under this protocol. Spatial, differential, and spectral diagnostics show broad terrain correspondence alongside smoothing, amplitude compression, and boundary mismatch. MarsFM provides a framework for combining learned terrain priors with image-based constraints; establishing improved physical terrain resolution requires matched baselines and independent high-resolution reference data. Data: https://huggingface.co/datasets/SuperComputer/mars_hirise_dtm_processed-6aa9b66ba461e07f; code: https://github.com/Marius-Juston/MarsRecon.