Invited Talks
A Single GPU Is All You Need for Self-Supervised Pretraining
Satellite Image Deep Learning Podcast, June 2026
I joined host Robin Cole to talk about self-supervised learning for remote sensing: predictive, generative, and contrastive approaches, how foundation models like DINO have reshaped computer vision, and why assumptions that hold for natural images often break down for satellite imagery. We then dig into my WACV 2026 workshop paper on subimage overlap prediction, which pretrains a model to locate a subimage within its source image as a proxy task for semantic segmentation. The method matches or beats established self-supervised approaches while pretraining on thousands of images rather than millions. We close on data efficiency, transfer across datasets, and why targeted pretraining is a practical alternative to ever-larger foundation models for niche geospatial applications.