Research map / Image generation / Autoregressive and latent generation
Revisiting IM2GPS in the Deep Learning Era
Nam Vo, Nathan Jacobs, 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 |
|---|---|---|---|
| use kNN kernel density retrieval instead of fine-grained classification | VGG-16 [L] 7011C · Im2GPS test set | accuracy within 1km: 6.8 -> 12.2 | Improved |
| use kNN retrieval (sigma=4) instead of PlaNet-style classification at continent level | VGG-16 [L] · Im2GPS test set | accuracy within 2500km: 71.3 -> 71.3 | No clear effect |
| replace nearest neighbour with kNN density estimation (sigma=1) (ablation) | [L] retrieval, Im2GPS database · Im2GPS test set | accuracy within 1km: 12.7 -> 13.1 | Improved |
| increase kernel bandwidth sigma from 4 to 16 (ablation) | [L] kNN, Im2GPS database · Im2GPS test set | accuracy within 1km: 12.2 -> 10.6 | Hurt |
| increase kernel bandwidth sigma from 4 to 16 (ablation) | [L] kNN, Im2GPS database · Im2GPS test set | accuracy within 2500km: 71.3 -> 75.9 | Improved |
| enlarge reference database from Im2GPS 6M to 28M images (ablation) | [L] kNN sigma=4 · Im2GPS test set | accuracy within 1km: 12.2 -> 14.4 | Improved |
| enlarge reference database from Im2GPS 6M to 28M images (ablation) | [L] kNN sigma=4 · Im2GPS test set | accuracy within 2500km: 71.3 -> 73.4 | Improved |
| replace Im2GPS database with YFCC22m database only (ablation) | [L] NN · Im2GPS test set | accuracy within 1km: 12.7 -> 12.2 | Hurt |
| add DML triplet ranking fine-tuning to the classification network (ablation) | VGG-16 [L] · Im2GPS3k and YFCC4k | geolocalization accuracy | No clear effect |
| add multiple classification losses (six partitions, model M) instead of single 7011-way loss (ablation) | VGG-16 [L] · Im2GPS3k and YFCC4k | geolocalization accuracy | Mixed |
| use finer partitioning of the Earth (7011 vs 80 or 359 classes) (ablation) | classification · Im2GPS3k and YFCC4k | accuracy at coarse and fine thresholds | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PlaNet [36] — PlaNet - Photo Geolocation with Convolutional Neural Networks (ECCV 2016) · better on 3 of 3 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
- Geolocation Estimation of Photos using a Hierarchical Model and Scene Classification (ECCV 2018 · similarity 0.89)
- GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.88)
- G3: An Effective and Adaptive Framework for Worldwide Geolocalization Using Large Multi-Modality Models (NeurIPS 2024 · similarity 0.86)
- PIGEON: Predicting Image Geolocations (CVPR 2024 · similarity 0.85)
- GeoRouter: Dynamic Paradigm Routing for Worldwide Image Geolocalization (NeurIPS 2026 · similarity 0.85)
- Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · 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.
