Computer Vision for Science 2026

Title: Active Measurement: Turning Computer Vision Models into Scientific Instruments  

Abstract:  Large observational datasets from radars, satellites, telescopes, cameras, and acoustic sensors are transforming the scale of scientific measurement. Computer vision models can help extract signals from these data, but scientific discovery often requires more than predictions: it requires calibrated measurements, uncertainty estimates, and efficient use of limited expert effort. In this talk, I will present active measurement, a human-in-the-loop framework that combines vision models, adaptive sampling, and statistical estimation to produce reliable measurements from large image collections. I will illustrate the approach through applications in ecology and biodiversity monitoring, including bird migration from weather radar, roost detection, and high-resolution bird counting, and discuss recent extensions to population-size estimation, multi-target measurement, and astronomical correlation functions. 

Title:  Imaging for Science

Abstract:  I'll describe my collaboration experiences with the Event Horizon Telescope collaboration that made the first image of a black hole. I'll provide details of the collaboration, and summarize with lessons we learned. I'll mention five application areas I feel are ripe for contributions from computer vision.  I'll close with a recipe for making contributions to scientific imaging. 

Title:  From Neural Rendering to Scientific Inference: 3D Reconstruction of Solar Eruptions with Physics-Informed NeRFs

Abstract:  Many scientific imaging problems require recovering three-dimensional structure from sparse, indirect, and projection-based measurements. Solar observations provide a particularly challenging example: the upper solar atmosphere and heliosphere are optically thin, dynamic, and observed from only a small number of viewpoints. This makes reconstructing the solar atmosphere intrinsically ill posed, yet its 3D structure is the missing link for understanding the Sun as it is: an evolving star embedded in a dynamic heliosphere.

Title:  Are remote sensing foundation models ready for science applications?

Abstract:  Foundation models are increasingly positioned as a transformative technology for science, with the promise of enabling generalizable knowledge across tasks, sensors, and domains. Early successes in remote sensing foundation models suggest great potential for Earth observation applications such as agriculture monitoring, biodiversity, and ecosystem research. But beneath the excitement lies a fundamental problem: as a field, we do not have a good understanding of the performance and robustness of remote sensing foundation models. This talk examines troubling trends in current foundation model research and suggests future directions toward building foundation models that are genuinely useful for scientific applications and discovery.

Title:  Multimodal Generative Models for Biomedical Discovery

Abstract:  This talk will work towards two directions of multimodal generative models that aim to assistant scientists in generating new biomedical insights. We will discuss image generative virtual cell models that can visually simulate the effect of genetic and drug perturbations, and VQA models that can aid in interpreting and analyzing experimental imaging data.