Research map / 3D vision / Depth and stereo
Co-Attention for Conditioned Image Matching
Olivia Wiles, Sebastien Ehrhardt, Andrew Zisserman
Check 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 CoAM conditioning to the CAPSNet descriptor model | CAPSNet · MegaDepth pose prediction, especially hard viewpoint pairs | correct rotational predictions (%): 65.8 -> 69.3 | Improved |
| Inject CoAM at both coarse and fine resolutions instead of coarse only (ablation) | CAPSNet · MegaDepth pose prediction | correct rotational predictions (%): easy: 91.7 -> 91.9; medium: 82.9 -> 82.8; hard: 69.3 -> 68.8 | Mixed |
| Add the distinctiveness score to matching (ablation) | CD-UNet · YFCC100M | mAP@5°: 31.60 -> 55.58 | Improved |
| Add CoAM conditioning (ablation) | CD-UNet · YFCC100M | mAP@5°: 53.43 -> 55.58 | Improved |
| Replace ResNet50 with EfficientNet-B1 (ablation) | CD-UNet · YFCC100M | mAP@5°: 55.58 -> 57.23 | Improved |
| Replace hinge loss with InfoNCE (ablation) | CD-UNet · Aachen Day-Night | accuracy at 0.5m (5°): 70.4 -> 62.2 | Hurt |
| Increase test-time grid size from G=128 to G=256 (ablation) | CD-UNet · Aachen Day-Night | accuracy at 0.5m (5°): 70.4 -> 68.4 | No clear effect |
| Use the local-neighborhood correspondence refinement strategy (ablation) | CD-UNet · HPatches | matching accuracy | Improved |
| Use EfficientNet-B1 instead of ResNet50 (ablation) | CD-UNet · HPatches | matching accuracy | Improved |
| Use pair-conditioned matching rather than SIFT features | 3D reconstruction pipeline · Large SfM: landmark datasets | registered images: 500 -> 702; 1035 -> 1072; 804 -> 967 | Improved |
| Use pair-conditioned matching rather than SIFT features | 3D reconstruction pipeline · Large SfM: landmark datasets | reprojection error (px): 0.60 -> 1.30; 0.69 -> 1.34; 0.61 -> 1.32 | Hurt |
| Use correspondence-based stylization instead of semantic stylization | Stylization model · MegaDepth | mean L1 error: 0.22 -> 0.14 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- RANSAC-Flow — RANSAC-Flow: Generic Two-stage Image Alignment (ECCV 2020) · better on 0 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Contrastive Attention Maps for Self-Supervised Co-Localization (ICCV 2021 · similarity 0.87)
- TransforMatcher: Match-to-Match Attention for Semantic Correspondence (CVPR 2022 · similarity 0.87)
- ASpanFormer: Detector-Free Image Matching with Adaptive Span Transformer (ECCV 2022 · similarity 0.86)
- ResMatch: Residual Attention Learning for Feature Matching (AAAI 2024 · similarity 0.85)
- Unifying Feature and Cost Aggregation with Transformers for Semantic and Visual Correspondence (ICLR 2024 · similarity 0.85)
- MatchDet: A Collaborative Framework for Image Matching and Object Detection (AAAI 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.
