Julia Romero, PhD

I recently completed my PhD in Computer Science at the University of Colorado Boulder. I build effective and efficient AI/ML approaches for spatiotemporal vision, multimodal, and sensor data, with applications in geospatial AI, video understanding, and ubiquitous computing. My work focuses on supervised and self-supervised learning, multimodal modeling, and practical approaches for real-world use. Specifically, my research has spanned these topics:

• Vision & Geospatial Foundation Models: I develop and interpret self-supervised vision foundation models for Earth observation tasks, examining pretraining, transfer learning, and downstream adaptation. For this work, I implemented and maintained multi-cluster distributed computing infrastructure on an NSF supercomputer [under review, 2026].
• Multimodal Video Understanding: I build lightweight and flexible graph-based vision models for fine-grained activity recognition. Our framework can be trained on multimodal and multi-view (multiple cameras) data while only requiring a single input for inference [ICCVW 2025]. This work with Intel Labs won 1st place out fo 20+ teams in the Ego-Exo4D Keystep Recognition challenge (sponsored by Meta) and led to an invited spotlight talk at the CVPR 2025 Egocentric Vision Workshop.
• Human-Centered Sensing Systems: I design sensing algorithms for resource-constrained wearables and ML pipelines for health and fitness data, spanning physiological monitoring with smart earbuds [HotMobile 2024], injury-risk modeling from wearables data, activity recognition with motion sensors [Sensors 2023], and large-scale physical-activity behavior analysis [IJERPH 2022].

My PhD research was funded in part by an NSF IUCRC award from a proposal I co-wrote, which supported a collaboration with Intel Labs from 2023 to 2025. During my PhD, I was also a research intern at Nokia Bell Labs in Cambridge, UK and worked on wearable computing and health ML at Stryd, Kinesis Integrated, and the Johns Hopkins Applied Physics Laboratory.