Research map / Label-efficient and robust learning / Knowledge distillation
Geometric Foundation Model Distillation for Efficient Lunar 3D Reconstruction
Clémentine Grethen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SVD-based decoder initialization from teacher weights (ablation) | S2 student · StereoLunar S2 training ablation | Chamfer: 0.68% -> 0.53% | Improved |
| add feature alignment loss (ablation) | S2 student · StereoLunar S2 training ablation | Chamfer: 0.67% -> 0.53% | Improved |
| unfreeze the DINOv2 encoder (ablation) | S2 student · StereoLunar S2 training ablation | Chamfer: 0.87% -> 0.53% | Improved |
| use teacher pseudo-labels instead of ground-truth supervision (ablation) | S2 student · StereoLunar S2 training ablation | Chamfer: 0.65% -> 0.53% | Improved |
| use ViT-Small encoder instead of MobileNet (ablation) | S2 vs S1 · StereoLunar test (1200 pairs) | Chamfer: 1.02% -> 0.53% | Improved |
| use ViT-Small instead of ViT-Tiny encoder (same decoder) (ablation) | S2 vs S4 · StereoLunar test (1200 pairs) | Chamfer: 0.74% -> 0.53% | Improved |
| shrink decoder from 512/6/4 to 256/4/4 (keep ViT-Small) (ablation) | S3 vs S2 · StereoLunar test (1200 pairs) | Chamfer: 0.53% -> 0.68% | Hurt |
| keep ViT-Small encoder with reduced decoder instead of ViT-Tiny with larger decoder (ablation) | S3 vs S4 · StereoLunar test (1200 pairs) | Chamfer: 0.74% -> 0.68% | Improved |
| add SVD init and feature alignment to DUNE student (ablation) | Dune S5 to S6 · StereoLunar encoder comparison (epoch 10) | Chamfer: 1.29% -> 1.21% | Improved |
| enlarge decoder to 512/6/4 for DUNE student (ablation) | Dune S6 to S7 · StereoLunar encoder comparison (epoch 10) | Chamfer: 1.21% -> 1.05% | Improved |
| use DINOv2 encoder instead of DUNE encoder (ablation) | S2 vs Dune S7 · StereoLunar encoder comparison (epoch 10) | Chamfer: 1.05% -> 0.90% | Improved |
| change layer mapping from uniform to first-k (ablation) | S2 decoder SVD init · StereoLunar layer mapping (25 epochs, subset) | Chamfer: 0.87% -> 0.85% | No clear effect |
Closest papers by meaning
- Mars-Bench: A Benchmark for Evaluating Foundation Models for Mars Science Tasks (NeurIPS 2025 · similarity 0.84)
- Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation (NeurIPS 2025 · similarity 0.84)
- LRM-Zero: Training Large Reconstruction Models with Synthesized Data (NeurIPS 2024 · similarity 0.83)
- Phaedra: Learning High-Fidelity Discrete Tokenization for the Physical Sciences (NeurIPS 2026 · similarity 0.83)
- LoRA3D: Low-Rank Self-Calibration of 3D Geometric Foundation models (ICLR 2025 · similarity 0.83)
- Distilling Monocular Foundation Model for Fine-grained Depth Completion (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
