#CVPR2025 Sat June 14 (PM)
🌍 Around the World in 80 Timesteps: A Generative Approach to Global Visual Geolocation
@nicolasdufour.bsky.social @vickykalogeiton.bsky.social @davidpicard.bsky.social @loicland.bsky.social
📄 pdf: arxiv.org/abs/2412.06781
🌐 webpage: nicolas-dufour.github.io/plonk.html

Very fine-grained geolocation may be better viewed a retrieval problem. We want to learn generalizable geographic features from images. In osv5m.github.io we added a 1km buffer between train and test to penalize overfitting. The train and test cars are also different to discourage "meta" learning.

Catch us at #NeurIPS2024 Dataset and Benchmarks! 🎉
🖼️ Poster: 5302 | 12/12 | 11 AM
📜 Paper: openreview.net/pdf?id=QpF3D...
🌐 Web & Data: archaeoscape.ai/data/2024
🤝 Joint Work: ENPC + French School of Asian Studies (#EFEO)

Archaeoscape has 2× the area and 3× the labels of comparable archaeological datasets, and is fully open-access! We performed an extensive benchmark of modern CV models and showed that segmenting archaeological traces is a surprisingly tough challenge—even for foundation models.

Khmer vestiges go far beyond monumental stone temples. Wooden and earthen structures faded centuries ago but left subtle geometric elevation patterns in the landscape. Can you spot these hidden structures on the LiDAR elevation maps? 🔍