Junwei Zhou

I'm a first year Ph.D. student at Dartmouth College, advised by Prof. Yu-Wing Tai. Previously, I have worked with Prof. Ming-Hsuan Yang, Prof. Lu Qi, Dr. Xueting Li at UC Merced, and Prof. Xinggang Wang at HUST. I received my B.E. from Huazhong University of Science and Technology.

My research focuses on fusing multimodal information to realize spatial intelligence that reconstruct, understand, and reason about the real world. Specifically, 3D vision & generative models, MLLMs for 3D spatial understanding.

Always actively seeking for research internship opportunities. Please drop me an email if you are interested.

Also, I am open to collaborations, please drop me an email if you want to have a chat.

Email  /  Google Scholar  /  Github

Research

We propose a perceive-then-plan framework with two VLMs (Perceiver and LaP Planner) for monocular 3D layout estimation, enabling both visual alignment and scene Coherence.

We introduce IDEAL-Bench, a benchmark designed to evaluate true spatial understanding by requiring models to infer structured 3D scene layouts from a single image, addressing the limitations of existing benchmarks.

Intergrating 2D amodal completion and 3D generative modeling to achieve amodal 3D objects generation from multiple partial occluded input images.

Generating a comprehensive and complex 4D scene by dividing a 4D scene and progressive dynamic content extrapolation.

Given a text prompt describing multiple objects and their spatial relationships, our method generates a 3D scene depicting these objects naturally interacting with one another.

SAM in helping weakly-supervised instance perception task.

Website borrowed from JonBarron.