Visualize OlmoEarth v1.2 Satellite Embeddings with FiftyOne

OlmoEarth v1.2, developed by Allen AI, is a foundation model for satellite imagery that significantly reduces computational costs without sacrificing accuracy, achieving 3× fewer GPU-hours during training and 2.9× fewer operations during inference compared to its predecessor, v1. This efficiency is achieved through advanced engineering techniques such as token collapsing and improved masking. It supports various satellite data including Sentinel-1, Sentinel-2, and Landsat, and can operate on Apple Silicon without a GPU. FiftyOne, Voxel51's open-source toolkit, complements OlmoEarth by visualizing and managing the model's high-dimensional embeddings, allowing users to explore satellite imagery through UMAP-reduced vectors. This integration enables unsupervised clustering and provides a synchronized view of embedding space, image grids, and geospatial maps. Both OlmoEarth and FiftyOne are open-source, with OlmoEarth's models and dataset openly available and including a license that prohibits military and extractive-industry applications.

No tracked trend matches for this post yet.