Research map / Architectures and efficiency / RNNs and normalization
MCNet: Rethinking the Core Ingredients for Accurate and Efficient Homography Estimation
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 MCNet, an iterative deep homography estimator that runs correlation searching across multiple feature scales with one-pass feature extraction, plus an FGO loss that boosts accuracy near convergence.
Problem
Deep homography estimation is accurate only with many iterations and feature re-extraction; single-scale correlation searching yields limited accuracy gains in later iterations, and dynamic-object scenes are hard, while prior methods are computationally expensive.
Approach
Combine iterative correlation searching with a coarse-to-fine multiscale feature pyramid extracted once (no per-iteration re-extraction or warping), and add a Fine-Grained Optimization loss, an inverse-proportional term on L1 error that increases gradients for well-estimated samples near convergence.
Key results
- On MSCOCO, MCNet reduces MACE by 20.5% vs 2-scale RHWF (0.031 vs 0.039) while cutting inference time by 78.9% (33.1 ms vs 157.1 ms) and FLOPs by 73.5%.
- On GoogleEarth and GoogleMap, MCNet cuts MACE by 43.4% and 41.1% relative to 2-scale RHWF.
- On SPID dynamic-object data, MCNet reduces MACE by 47.1% relative to 2-scale IHN-mov (0.680 vs 1.285).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MCNet instead of 2-scale RHWF | homography estimation · MSCOCO test | MACE: 0.039 -> 0.031 | Helped |
| use MCNet instead of 2-scale RHWF | homography estimation · GoogleEarth | MACE: 1.066 -> 0.603 | Helped |
| use MCNet instead of 2-scale RHWF | homography estimation · GoogleMap | MACE: 0.394 -> 0.232 | Helped |
| use MCNet instead of 2-scale IHN-mov | homography estimation · SPID | MACE: 1.285 -> 0.680 | Helped |
| increase iterations per scale from (2, 2, 2) to (4, 4, 4) (ablation) | MCNet · MSCOCO | MACE: 0.031 -> 0.023 | Helped |
| increase iterations per scale from (2, 2, 2) to (8, 8, 8) (ablation) | MCNet · MSCOCO | MACE: 0.031 -> 0.022 | Helped |
| increase iterations per scale to (4, 4, 4) (ablation) | MCNet · MSCOCO | inference time: 33.1 -> 55.3 | Hurt |
| increase iterations per scale to (8, 8, 8) (ablation) | MCNet · MSCOCO | inference time: 33.1 -> 101.0 | Hurt |
| use 2 scales instead of 1 (ablation) | MCNet · MSCOCO | MACE: 0.336 -> 0.072 | Helped |
| use 3 scales instead of 1 (ablation) | MCNet · MSCOCO | MACE: 0.336 -> 0.031 | Helped |
| add FGO loss to L1 loss (ϵ=0.1, α=0.85) (ablation) | MCNet · MSCOCO | MACE: 0.046 -> 0.031 | Helped |
| set FGO parameters to ϵ=0.001, α=1 (ablation) | MCNet with L1+FGO · MSCOCO | MACE: 0.046 -> 0.038 | Helped |
Surprising findings
- Adding a higher-resolution scale to correlation searching improves accuracy with negligible extra time, and MCNet handles dynamic foreground objects without explicit mask generation.
Limitations
- Evaluation limited to 128x128 image pairs with synthetic perturbations on MSCOCO.
- FGO loss hyperparameters (epsilon, alpha) were tuned via ablation on MSCOCO.
Experimental setup
Datasets: MSCOCO GoogleEarth GoogleMap SPID
Benchmarks: MSCOCO GoogleEarth GoogleMap SPID
Metrics: average corner error (ACE) mean average corner error (MACE) inference time FLOPs parameters memory
Baselines: RHWF IHN LocalTrans MHN UDHN DHN CLKN DLKFM SIFT+RANSAC SIFT+MAGSAC
Builds on: IHN RHWF DeTone deep homography (DHN) IC-LK iterator
Closest papers by meaning
- Adapting Dense Matching for Homography Estimation with Grid-based Acceleration (CVPR 2025 · similarity 0.92)
- SSHNet: Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization (CVPR 2025 · similarity 0.90)
- Noisy One-point Homographies are Surprisingly Good (CVPR 2024 · similarity 0.87)
- Unsupervised Homography Estimation on Multimodal Image Pair via Alternating Optimization (NeurIPS 2024 · similarity 0.86)
- Mask-Homo: Pseudo Plane Mask-Guided Unsupervised Multi-Homography Estimation (AAAI 2024 · similarity 0.86)
- HUNet: Homotopy Unfolding Network for Image Compressive Sensing (CVPR 2025 · similarity 0.85)
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.