IEEE European Conference on Computer Vision (ECCV 2026)

Geometric Foundation Model Distillation
for Efficient Lunar 3D Reconstruction

Clémentine Grethen¹, Géraldine Morin^1,*, Simone Gasparini^1,*, Florient Chouteau^2,*

¹IRIT, University of Toulouse, France · ²Airbus Defence and Space

* Authors omitted from the official ECCV record due to a submission error that the editorial process did not allow us to correct after acceptance. Their contributions are gratefully acknowledged here.

GitHub Paper

Abstract

Large geometric foundation models (e.g. MASt3R, 688.6M parameters) achieve state-of-the-art 3D reconstruction from stereo pairs but are prohibitive to deploy on resource-constrained platforms — particularly relevant for lunar space missions. We propose a distillation framework combining (i) SVD-based decoder initialization from the teacher (Eckart–Young optimal), (ii) a feature-alignment loss at intermediate ViT layers, and (iii) full encoder fine-tuning.

Method

Teacher MASt3R is distilled into a lighter student with SVD decoder initialization, feature alignment, and full encoder fine-tuning.

Choose a stereo pair

Inference was run with our actual checkpoint-50 for every model on each pair. Pick one — all comparisons below update accordingly.

Student Architecture Performance Ablation

S2 BEST uses 154.9M parameters versus 688.6M for Teacher = MASt3R. The reduction is the point.

Depth map

Teacher vs selected student · View pair: left / right

Teacher

Student

Point Cloud

Student and Teacher

Student

Selected architecture

Teacher

MASt3R · 688.6M

Distillation Strategy Ablation

Paper ablation study on S2

Compare BEST against no SVD / frozen encoder / no feature loss / GT supervision. Depth and slope both available.

Best:

SVD Frozen enc. Feat. GT

Selected: B

SVD Frozen enc. Feat. GT

Distillation ablation

A (BEST) vs selected variant · depth / slope

A (BEST)

Citation

If you use this work, please cite the paper below.

@inproceedings{grethen2026geometric, title={Geometric Foundation Model Distillation for Efficient Lunar 3D Reconstruction}, author={Grethen, Clémentine and Morin, Géraldine and Gasparini, Simone and Chouteau, Florient}, booktitle={European Conference on Computer Vision (ECCV)}, year={2026} }

Acknowledgements

This work was supported by the French Agence Nationale de la Recherche (ANR, “Investissements d’avenir”, ANR-21-ESRE-0051) and the European Space Agency (ESA, contract 4000140461/23/NL/GLC/my).