Research map / Label-efficient and robust learning / Recognition and long-tail
Deep Homography Estimation for Visual Place Recognition
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 DHE-VPR, a hierarchical visual place recognition method where a transformer-based deep homography network replaces RANSAC for geometric verification, trained with a re-projection error of inliers loss jointly with the backbone.
Problem
Hierarchical visual place recognition re-ranking relies on RANSAC homography fitting, which is slow and non-differentiable, so networks are usually trained only for global features.
Approach
Regress a homography from a similarity map of dense local features with a transformer (no homography labels), using RANSAC inlier counts as weak supervision via a re-projection error of inliers loss that back-propagates into the backbone; inlier count of the learned homography is the re-ranking score.
Key results
- DHE-VPR averages R@1 80.3, R@5 89.9, R@10 92.1 across Pitts30k, MSLS val/challenge, Nordland and St. Lucia, the best average among compared methods.
- On Nordland, DHE-VPR reaches R@1 67.4 versus 46.8 for its own GeM-only retrieval, a 20.6-point gain from re-ranking.
- Matching time on Pitts30k is 0.098 s per query versus 3.010 s for TransVPR and 10.732 s for Patch-NetVLAD-p, over an order of magnitude faster than RANSAC-based re-ranking.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DHE geometric re-ranking to GeM retrieval without joint fine-tuning (ablation) | DHE-VPR · Pitts30k | R@1: 83.1 -> 87.8 | Helped |
| Add DHE geometric re-ranking to GeM retrieval without joint fine-tuning (ablation) | DHE-VPR · MSLS val | R@1: 80.5 -> 81.2 | Helped |
| Jointly fine-tune the backbone and DHE with the proposed training strategy, without re-ranking (ablation) | DHE-VPR global retrieval · Pitts30k | R@1: 83.1 -> 83.9 | No clear effect |
| Jointly fine-tune the backbone and DHE with the proposed training strategy, without re-ranking (ablation) | DHE-VPR global retrieval · MSLS val | R@1: 80.5 -> 80.3 | No clear effect |
| Add joint fine-tuning with the proposed training strategy to DHE re-ranking (ablation) | DHE-VPR · Pitts30k | R@1: 87.8 -> 89.4 | Helped |
| Add joint fine-tuning with the proposed training strategy to DHE re-ranking (ablation) | DHE-VPR · MSLS val | R@1: 81.2 -> 84.1 | Helped |
| Use DHE geometric re-ranking rather than RANSAC geometric re-ranking (ablation) | DHE-VPR · Pitts30k | R@1: 88.3 -> 89.4 | Helped |
| Use geometric-verification DHE rather than image-alignment GeoWarp (ablation) | VGG16 re-ranking · Pitts30k | R@1: 65.3 -> 67.6 | Helped |
| Use geometric-verification DHE rather than image-alignment GeoWarp (ablation) | VGG16 re-ranking · MSLS val | R@1: 48.0 -> 51.9 | Helped |
| Use DHE-VPR rather than TransVPR | VPR · MSLS val | R@1: 86.8 -> 84.1 | Hurt |
| Use DHE-VPR rather than CosPlace | VPR · Nordland test | R@1: 64.1 -> 67.4 | Helped |
| Use DHE-VPR rather than TransVPR | VPR · St. Lucia | R@1: 98.7 -> 99.1 | Helped |
Surprising findings
- Fine-tuning with the REI loss, which uses only RANSAC inlier counts as supervision, lets DHE-VPR outperform the RANSAC-based re-ranking pipeline built on the same backbone.
Limitations
- Uses a single homography per image pair, which the authors motivate for planar or rotational cases but which is only approximate for general VPR scenes.
- Relies on RANSAC inlier counts as supervision, so training still depends on RANSAC during the training stage.
- Compares at 384x384 resolution while baselines use 640x480, making comparisons not fully matched.
Experimental setup
Datasets: MSLS Pitts30k Nordland St. Lucia ImageNet
Benchmarks: Pitts30k MSLS val MSLS challenge Nordland St. Lucia
Metrics: Recall@1 Recall@5 Recall@10 runtime (s)
Baselines: NetVLAD SFRS CosPlace SP-SuperGlue Patch-NetVLAD TransVPR ETR-D GeoWarp
Builds on: CCT NetVLAD GeoWarp RANSAC Direct Linear Transform
Closest papers by meaning
- Towards Seamless Adaptation of Pre-trained Models for Visual Place Recognition (ICLR 2024 · similarity 0.86)
- Adapting Dense Matching for Homography Estimation with Grid-based Acceleration (CVPR 2025 · similarity 0.86)
- Towards Implicit Aggregation: Robust Image Representation for Place Recognition in the Transformer Era (NeurIPS 2025 · similarity 0.86)
- SuperVLAD: Compact and Robust Image Descriptors for Visual Place Recognition (NeurIPS 2024 · similarity 0.86)
- Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · similarity 0.85)
- MCNet: Rethinking the Core Ingredients for Accurate and Efficient Homography Estimation (CVPR 2024 · 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.