Yuqun Wu

I am a final-year Ph.D. student in Computer Science at the University of Illinois at Urbana-Champaign, advised by Prof. Derek Hoiem and Prof. Shenlong Wang.

I received my MS and BS degrees from UIUC, where I was fortunate to be mentored by Dr. Jae Yong Lee. I have interned at Stability AI (2026) and Meta (2025).

Email  /  CV  /  Linkedin

My research interests lie in computer vision, with a current focus on physically grounded world models and VLMs for spatial understanding. I am also broadly interested in scene understanding, neural rendering and geometry reconstruction.

Propose novel-view semantic rendering as an auxiliary training task that helps VLMs integrate geometry features for better spatial reasoning.

Integrate VGGT and SAM for consistent multi-view instance segmentation, without offline mask matching or explicit 3D reconstruction.

Propose a learnable retrieval token (ReToken) that retrieves the relevant frames and improves the VLM performance.

Propose a dataset and a new method for object change detection between a pair of captures (images or videos) of the same scene at different times.

Propose a training-free approach to create text compatible region tokens, enabling powerful zero-shot region-level understanding with existing image-text models.

Create 3D models that provide accurate geometry and view synthesis, partially closing the large geometric performance gap between NeRF and traditional MVS methods

Encode 3D scenes into extremely compact representation from 2D images and enable its transmittance, decoding and rendering in real-time across various platforms via traditional GL pipeline.

Investigate region based representation, combining class-agnostic segmentation from SAM and dense features from foundation models, for a wide variety of tasks, including semantic segmentation, object-based image retrieval, and multi-image analysis.

Propose a novel method that outperforms existing depth completion pipelines given sparse keypoint depth, and reconstructs complete point clouds given SfM setups

Present Quantized Fourier Features (QFF), which encodes features in bins of Fourier features, and can result in smaller model size, faster training, and better quality outputs for various applications of neural representation

Reviewer: CVPR, ECCV, 3DV, TMLR, WACV

Teaching Assistant: Applied Machine Learning(Fall 2026, Spring 2023), Computational Photography(Fall 2022)