Research map / 3D vision / Depth and stereo
Mask-Homo: Pseudo Plane Mask-Guided Unsupervised Multi-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 Mask-Homo, an unsupervised multi-homography estimator guided by semantic pseudo-plane masks, plus a Mask-PSNR metric for alignment evaluation.
Problem
Single-homography estimation compromises across multiple planes at different depths, while mesh-grid multi-homography methods break the scene's plane distribution and do not address single-H model inadequacy.
Approach
Derive correlated pseudo-plane mask pairs from semantic segmentation (with fragment removal and category/size/center matching), then predict one local homography per mask pair with a transformer using masked triplet and feature identity losses, and fuse the warped images.
Key results
- Multi-H result: 0.69 average MSE over all points vs 0.79 for MeshBasesHomo 2x2 mesh (about 13% improvement) using at most 4 homographies.
- Mask-PSNR average 41.93 dB on dominant-plane evaluation, outperforming HomoGAN* (f.t.) at 40.95 dB and the best existing method by 0.89 dB.
- Single-H dominant-plane MSE 0.41 vs HomoGAN* (p.t.) 0.49, about 16% improvement.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Mask-Homo instead of HomoGAN* (f.t.) | Multi-homography estimation · Natural image test set, dominant plane | MSE: 0.42→0.41 | Helped |
| Use Mask-Homo instead of MeshBasesHomo | Multi-homography estimation · Natural image test set, all points | MSE: 0.79→0.69 | Helped |
| Use Mask-Homo instead of existing methods | Multi-homography estimation · Natural image test set | Mask-PSNR: +0.89dB | Helped |
| Use Mask-Homo instead of HomoGAN* (f.t.) | Multi-homography estimation · LT, dominant plane | MSE: 0.40→0.61 | Hurt |
| Use Mask-Homo instead of RAFT | Image alignment · Natural image test set, all points (Avg) | MSE: 0.68 vs 0.69 | Mixed |
| Vary Dmin from 15 to 45 (ablation) | Mask-Homo pseudo plane mask generation · Natural image test set | MSE | Helped |
| Increase the maximum number of pseudo plane masks from 4 to 8 (ablation) | Mask-Homo pseudo plane mask generation · Natural image test set | MSE | Helped |
Surprising findings
- Mask-PSNR is claimed to align better with human observation than MSE on sparse matching points, as illustrated in Fig 2.
- Mask-Homo is slightly worse than baselines on the LT scene in dominant-plane MSE and clearly worse on LF scenes in all-points MSE, attributed to moving objects.
Limitations
- Mask-Homo performs worse on large-foreground (LF) scenes with moving objects, and slightly worse on low-texture scenes.
- Pseudo-plane masks come from semantic segmentation and heuristic post-processing with several hyperparameters (S_hole, S_seg, D_min, S_rto, N).
- MeshBasesHomo code unavailable, so it could not be evaluated with Mask-PSNR or qualitatively.
Experimental setup
Datasets: natural image dataset (Zhang et al. 2020; Liu et al. 2022a)
Benchmarks: natural image homography dataset (RE, LT, LL, SF, LF scenes)
Metrics: MSE Mask-PSNR
Baselines: Supervised (DeTone et al.) Unsupervised (Nguyen et al.) CA-Unsupervised BasesHomo HomoGAN APAP MeshFlow MeshBasesHomo PWCNet RAFT
Builds on: HomoGAN Content-Aware Unsupervised Deep Homography Estimation Motion Basis Learning (BasesHomo) APAP
Closest papers by meaning
- Noisy One-point Homographies are Surprisingly Good (CVPR 2024 · similarity 0.88)
- Adapting Dense Matching for Homography Estimation with Grid-based Acceleration (CVPR 2025 · similarity 0.87)
- MCNet: Rethinking the Core Ingredients for Accurate and Efficient Homography Estimation (CVPR 2024 · similarity 0.86)
- Unsupervised Homography Estimation on Multimodal Image Pair via Alternating Optimization (NeurIPS 2024 · similarity 0.85)
- SSHNet: Unsupervised Cross-modal Homography Estimation via Problem Reformulation and Split Optimization (CVPR 2025 · similarity 0.84)
- HomoGen: Enhanced Video Inpainting via Homography Propagation and Diffusion (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.