Ziqi Pang (庞子奇)

I am a fourth-year CS Ph.D. student focusing on computer vision and machine learning at University of Illinois Urabana-Champaign (UIUC), where my advisor is Prof. Yu-Xiong Wang. Before that, I graduated from Peking University (PKU) with a Bachelor degree in Computer Science.

I interned at Toyota Research Institute (TRI) with Dr. Pavel Tokmakov during my Ph.D. study. Prior to joining UIUC, I interned at Carnegie Mellon University (CMU) with Prof. Martial Hebert, practiced research at Peking University (PKU) with Prof. Shiliang Zhang, and spent an exciting year at TuSimple pushing the boundaries of autonomous driving guided by Dr. Naiyan Wang.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github  / 

Enhance the Knowledge in Generative Foundation Models for Embodied Perception in Long Videos. I care about embodied perception in 2D, 3D and 4D. I envision generative pre-trained model as the critical component enabling their scaling and self-improvement.

I am actively seeking summer 2025 research internships. Feel free to contact with me if you think I could make a contribution to your team.

We enable a GPT-style causal transformer to generate images in random orders, which unlocks a series of new capabilities for decoder-only autoregressive models.

We propose a simple yet effective MLLMs for language-instructed video segmentation. It emphasizes global-local video understanding and achieves SOTA performance on multiple benchmarks.

Our paper answers several critical question on diffusion models for visual perception: (1) how to train diffusion-based perception models, (2) how to utilize diffusion models as a unique interactive user interface.

Using the relationships between entities, we can better control the generation of images with multi-modal graphs.

Managing memory banks better significantly improves VOS on challenging state-changes and long videos. Similar strategy is also adopted in SAM2 later.

Frozen transformers from language models, though trained solely on textual data, can effectively improves diverse visual tasks by directly encoding visual tokens. It is an essential step for my research on "generative models benefiting perception."

MV-Map is the first offboard auto-labeling pipeline for HD-Maps, whose crust is to fuse BEV perception results guided by geometric cues from NeRFs.

"Streaming forecasting" mitigates the gap between "snapshot-based" conventional motion forecasting and the streaming real-world traffic.

PF-Track is an vision-centric 3D MOT framework that dramatically decreases ID-Switches by 90% with an end-to-end framework for autonomous driving.

SST emphasize the small object sizes and sparsity of point clouds. Its sparse transformers enlight new backbones for outdoor LiDAR-based detection.

SimpleTrack is simple-yet-effective 3D MOT system. It is one of the most widely adopted 3D MOT baseline worldwide.

LiDAR-SOT is a LiDAR-based data flywheeel and auto-labeling pipeline for autonomous driving.

Huge thanks to Jon Barron for proving the template for the page.