Research map / 3D vision / Depth and stereo
Adapting Dense Matching for Homography Estimation with Grid-based Acceleration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GFNet, which adapts RoMa-style dense matching to homography estimation by regressing flow and confidence on a coarse grid initialized from DINOv2 patch matches, enabling high-resolution inputs and lower compute.
Problem
Deep homography estimation is limited to low-resolution inputs (about 128x128) and dense matching is computationally heavy for homography, which only has 8 degrees of freedom.
Approach
Replace pixel-wise dense flow with a grid flow regressed coarse-to-fine from DINOv2 patch matches plus a lightweight FPN, using iterative regression per scale, composite homographies for training symmetry, and adaptive resolution recurrence at inference; homography is then fit by RANSAC on matches from grid flow.
Key results
- On 448x448 GoogleMap, GFNet reaches 51.44 auc@3 vs RoMa 45.29 (+13.5%), with about 33.7% fewer MACs than RoMa without iteration.
- On MSCOCO at 448x448, GFNet achieves 97.69 auc@3, above RoMa's 96.01 (+1.7%) and RHWF's 93.02 (+5.0%).
- GFNet gives a 1.8x auc@3 improvement over the SOTA deep homography method on GoogleMap (51.44 vs 18.32 for RHWF).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GFNet instead of RHWF | homography estimation · MSCOCO | auc@3: 93.02 -> 97.69 | Helped |
| use GFNet instead of RoMa | homography estimation · VIRAT | auc@3: 47.5 -> 46.92 | Hurt |
| use GFNet instead of RoMa | homography estimation · VIS-IR | auc@3: 25.74 -> 21.28 | Hurt |
| use GFNet instead of RoMa | homography estimation · GoogleMap | auc@3: 45.29 -> 51.44 | Helped |
| use the proposed grid sizes instead of Half grid sizes (ablation) | GFNet · MSCOCO | MACE (pixel): 0.068 -> 0.063 | Helped |
| use the proposed grid sizes instead of Full grid sizes (ablation) | GFNet · GoogleMap | MACE (pixel): 4.02 -> 2.57 | Helped |
| increase iterations per scale from N=1 to N=2 (ablation) | GFNet · MSCOCO | MACE (pixel): 0.063 -> 0.135 | Hurt |
| increase iterations per scale from N=1 to N=2 (ablation) | GFNet · GoogleMap | MACE (pixel): 3.01 -> 2.57 | Helped |
| increase iterations per scale from N=2 to N=4 (ablation) | GFNet · GoogleMap | MACE (pixel): 2.57 -> 2.62 | Hurt |
| replace DINOv2 with a trainable pre-trained ResNet50 (ablation) | GFNet · GoogleMap | MACE (pixel): 2.57 -> 8.35 | Hurt |
| remove cross-attention layers (ablation) | GFNet · MSCOCO | MACE (pixel): 0.063 -> 0.071 | Hurt |
| remove symmetric inference (ablation) | GFNet · GoogleMap | MACE (pixel): 2.57 -> 2.95 | Hurt |
Surprising findings
- Iterative regression helps multimodal data (GoogleMap) but hurts natural-image accuracy on MSCOCO (MACE rises from 0.063 to 0.135 pixels with N=2).
- GFNet is worse than RoMa on VIS-IR (21.28 vs 25.74 auc@3), attributed to low-light degradation.
Limitations
- Iterative regression is only used for multimodal cases since it reduces accuracy on natural images.
- Performs worse than RoMa on VIS-IR low-light images.
- VIRAT test set is small (144 samples).
Experimental setup
Datasets: CityScapes ADE-20K GoogleMap VIS-IR MSCOCO VIRAT
Benchmarks: MSCOCO VIS-IR GoogleMap VIRAT
Metrics: auc@3 auc@5 auc@10 auc@20 ACE MACs runtime
Baselines: RHWF MCNet PRISE SIFT+LightGlue LoFTR Efficient LoFTR RoMa
Builds on: RoMa DINOv2 IHN RAFT PRISE
Closest papers by meaning
- MCNet: Rethinking the Core Ingredients for Accurate and Efficient Homography Estimation (CVPR 2024 · similarity 0.92)
- HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation (AAAI 2025 · similarity 0.90)
- Noisy One-point Homographies are Surprisingly Good (CVPR 2024 · similarity 0.88)
- SSHNet: Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization (CVPR 2025 · similarity 0.87)
- UFM: A Simple Path towards Unified Dense Correspondence with Flow (NeurIPS 2025 · similarity 0.87)
- Mask-Homo: Pseudo Plane Mask-Guided Unsupervised Multi-Homography Estimation (AAAI 2024 · similarity 0.87)
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.