PlaNet CNN geolocates photos, claims superhuman accuracy
TL;DR
- PlaNet reframes photo geolocation as classification over thousands of multi-scale geographic cells rather than image retrieval.
- The convolutional network is trained on millions of geotagged images and integrates cues like weather, vegetation and architecture.
- Combining the model with an LSTM over full photo albums yields a 50% performance improvement over single-image inference.
PlaNet, a convolutional network from researchers Tobias Weyand, Ilya Kostrikov and James Philbin, treats photo geolocation as a classification problem rather than an image-retrieval one, dividing the planet into thousands of multi-scale geographic cells and training on millions of geotagged images.
The system leans on cues the authors list as "landmarks, weather patterns, vegetation, road markings, and architectural details." They motivate the setup by way of the game GeoGuessr and the blog View from your Window, arguing humans are "relatively good at integrating these cues to geolocate images, especially en-masse."
The paper reports that PlaNet "outperforms previous approaches and even attains superhuman levels of accuracy in some cases," though the abstract publishes no per-scale accuracy numbers and does not name the baselines it beats. Extended with a long short-term memory architecture to reason over full photo albums, the same model "achieves a 50% performance improvement over the single-image model" by exploiting temporal coherence across a trip's images.
Shared on Bluesky by 1 AI expert
Originally reported by arxiv.org
Read the original article →Original headline: PlaNet - Photo Geolocation with Convolutional Neural Networks