PIGEON: Predicting Image Geolocations

TL;DR

This work tackles planet-scale image geolocalization by combining semantic geocell construction, multi-task contrastive pretraining with geographic auxiliary data, and a distance-aware, haversine-based smoothing loss. It introduces PIGEON (Street View panorama input) and PIGEOTTO (general image input), achieving strong results on Street View data and broad benchmark datasets, with PIGEOTTO showing robust generalization to unseen places. A hierarchical refinement via location-cluster retrieval further improves accuracy, enabling precise intra-cluster predictions. The approach advances state-of-the-art in geolocalization while highlighting ethical considerations and the need for responsible deployment of location-aware vision systems.

Abstract

Planet-scale image geolocalization remains a challenging problem due to the diversity of images originating from anywhere in the world. Although approaches based on vision transformers have made significant progress in geolocalization accuracy, success in prior literature is constrained to narrow distributions of images of landmarks, and performance has not generalized to unseen places. We present a new geolocalization system that combines semantic geocell creation, multi-task contrastive pretraining, and a novel loss function. Additionally, our work is the first to perform retrieval over location clusters for guess refinements. We train two models for evaluations on street-level data and general-purpose image geolocalization; the first model, PIGEON, is trained on data from the game of Geoguessr and is capable of placing over 40% of its guesses within 25 kilometers of the target location globally. We also develop a bot and deploy PIGEON in a blind experiment against humans, ranking in the top 0.01% of players. We further challenge one of the world's foremost professional Geoguessr players to a series of six matches with millions of viewers, winning all six games. Our second model, PIGEOTTO, differs in that it is trained on a dataset of images from Flickr and Wikipedia, achieving state-of-the-art results on a wide range of image geolocalization benchmarks, outperforming the previous SOTA by up to 7.7 percentage points on the city accuracy level and up to 38.8 percentage points on the country level. Our findings suggest that PIGEOTTO is the first image geolocalization model that effectively generalizes to unseen places and that our approach can pave the way for highly accurate, planet-scale image geolocalization systems. Our code is available on GitHub.

Figures (13)

• Figure 1: Prediction pipeline and main contributions of PIGEON. Administrative boundary and training set metadata are hierarchically ranked, clustered, and Voronoi tessellated to create semantic geocells. The geocell labels are then used to create continuous labels via haversine smoothing. Additionally, we pretrain CLIP via geographic synthetic captions in a multi-task setting. The pretrained CLIP model together with an OPTICS clustering model are employed to generate location cluster representations. During inference, an image embedding is computed and first passed to a linear layer to create geocell predictions and to identify the topK geocell candidates. The embedding is also used in our refinement pipeline to refine predictions within and across geocells. This is achieved by minimizing the embedding $L_2$-distance between the inference image embedding and all location cluster representations across the topK geocells. Finally, predictions are refined within the top identified cluster to generate geographic coordinates as outputs.
• Figure 2: Geocell specifications around Paris, France.
• Figure 3: Impact of applying haversine smoothing over neighboring geocells for a location in Accra, Ghana.
• Figure 4: Geolocalization error of PIGEON against human players of various in-game skill levels across 458 multi-round matches. The Champion Division consists of the top 0.01% of players. PIGEON's error is higher than in \ref{['table:ablation_study']} because GeoGuessr round difficulties are adjusted dynamically, increasing with every round.
• Figure 5: Voronoi tessellation applied in the process of geocell creation for points of an OPTICS cluster in Vienna, Austria, based on political boundaries from GADM gadm_2022.