Research map / Image generation / Autoregressive and latent generation
Localizing and Orienting Street Views Using Overhead Imagery
Nam N. Vo, James Hays
Read the paperCheck 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 |
|---|---|---|---|
| replace contrastive loss with distance-based logistic (DBL) layer in Siamese network (ablation) | Siamese network · Denver test set | ranking recall @top 1%: 21.6 -> 48.4 | Improved |
| replace hinge triplet loss with DBL layer (ablation) | Triplet network · Denver test set | ranking recall @top 1%: 43.2 -> 49.3 | Improved |
| use triplet instead of Siamese contrastive network (ablation) | cross-view embedding · Denver test set | ranking recall @top 1%: 21.6 -> 43.2 | Improved |
| process both images jointly in hybrid classification network instead of pure classification network (ablation) | classification network · Denver test set | classification accuracy: 90.0 -> 91.5 | Improved |
| increase rotation invariance from 0 to 90 degrees with 16 test rotations (ablation) | triplet DBL-Net · Denver test set | ranking recall @top 1%: 37.3 -> 49.3 | Improved |
| full 360 degree rotation invariance instead of 90 degrees at 16 test rotations (ablation) | triplet DBL-Net · Denver test set | ranking recall @top 1%: 49.3 -> 41.9 | Hurt |
| add orientation regression to 360 degree RI network (ablation) | triplet DBL-Net · Denver test set | ranking recall @top 1%: 41.9 -> 54.8 | Improved |
| add orientation regression to 90 degree RI network (ablation) | triplet DBL-Net · Denver test set | ranking recall @top 1%: 49.3 -> 47.4 | Hurt |
| increase test rotation samples from 1 to 16 for 0 degree RI network (ablation) | triplet DBL-Net · Denver test set | ranking recall @top 1%: 11.0 -> 37.3 | Improved |
| average features of 16 rotation samples at indexing time instead of testing all 16 (ablation) | 360 RI + OR network · Denver test set | ranking recall @top 1%: 54.8 -> 54.0 | No clear effect |
| use all triplets in mini-batch (eDBL) instead of single triplets (ablation) | triplet DBL-Net 360 RI + OR + avg16 · Denver test set | ranking recall @top 1%: 54.0 -> 59.9 | Improved |
Closest papers by meaning
- Lending Orientation to Neural Networks for Cross-View Geo-Localization (CVPR 2019 · similarity 0.91)
- Cross-View Image Matching for Geo-Localization in Urban Environments (CVPR 2017 · similarity 0.90)
- Coming Down to Earth: Satellite-to-Street View Synthesis for Geo-Localization (CVPR 2021 · similarity 0.88)
- CVM-Net: Cross-View Matching Network for Image-Based Ground-to-Aerial Geo-Localization (CVPR 2018 · similarity 0.88)
- OrienterNet: Visual Localization in 2D Public Maps With Neural Matching (CVPR 2023 · similarity 0.87)
- Cross-view Geo-localization with Layer-to-Layer Transformer (NeurIPS 2021 · 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.
