Jaewoo (Jeffrey) Heo

Hey! I'm Jaewoo (Jeffrey), an ML researcher at the Stanford MARVL (Medical AI and ComputeR Vision Lab) where I am advised by Professor Serena Yeung-Levy. I am also a part of Stanford Vision and Learning Lab, where I work with Professor Fei-Fei Li and Professor Ehsan Adeli in the Partner in AI-Assisted Care (PAC) group.

I received my B.S. in Computer Science with Honors from Stanford University in 2024, and my M.S in Computer Science from Stanford University in 2025. Throughout my career in ML, I've been primarily involved with computer vision and medical AI research. My recent works in query-agnostic deformable cross attention, score distillation sampling with diffusion models, long-form video understanding, and VLM for reliable data generation aim to develop machines with enhanced spatiotemporal understanding of our world.

In my free time, I enjoy playing the guitar, writing music, playing tennis, and watching soccer. I am a huge fan of The Beatles, Pink Floyd, and Billy Joel. I used to never be able to listen to music while coding, but I've recently developed a knack for putting on some bossa nova while working.

Email  /  CV  /  Scholar  /  Github

Research

"Understanding the problem is half the solution"

My research interests lie in computer vision and deep learning for perception and spatiotemporal understanding to make meaningful inferences of the world. Specifically, I'm passionate about applied ML research particularly in the fields of robotics and autonomous vehicles. My work aims to bridge the gap between technical innovation and real-world application.

Evaluated 11 state-of-the-art VLMs across 17 surgical AI tasks using 13 datasets, demonstrating VLMs' superior generalizability compared to supervised models when deployed outside their training environments

Proposes a novel query-agnostic deformable cross-attention mechanism that allows the model to attend to relevant spatial features more flexibly and in a data-dependent manner. Achieves SOTA performance on 3D human mesh recovery benchmarks 3DPW and RICH.

A 3D global HMR model that leverages the motion diffusion model (MDM) as a prior of coherent human motion. The model is robust to dynamic camera motion and long videos.

Proposes a novel data generation method for close interactions that leverages noisy automatic annotations to scale data acquisition, producing pseudo-ground truth meshes from in-the-wild images.

Rethinks the dependency on the 2D key point fitting paradigm and presents NeuHMR, an optimization-based mesh recovery framework based on recent advances in neural rendering (NeRF).

I referred to this website's source code to build my own.