Zihui Gao

I'm a Research Fellow at Nanyang Technological University (NTU), Singapore, working with Jiawang Bian. I work on 3D vision: recovering the geometry and appearance of the real world from images and videos, and turning it into representations that machines can perceive, simulate, and interact with.

I did my PhD at Zhejiang University, in the State Key Lab of CAD & CG, where I was advised by Hao Chen and Chunhua Shen.

Email  /  Scholar

Research

I'm interested in 3D computer vision, world models, and embodied AI. My work spans 3D reconstruction (sparse-view surface reconstruction and Gaussian splatting, and scaling 3D geometric foundation models with Internet video), 3D perception for autonomous driving (monocular 3D object detection and vision-language driving), and interactive video world models together with the large-scale data that trains them.

Going forward, I'm particularly excited about real-to-sim: reconstructing real-world scenes and objects into faithful, interactive simulations, so that robots and embodied agents can be trained and evaluated in digital twins of the world they will act in.

A large-scale egocentric video dataset pairing long videos with camera trajectories and temporally grounded captions, for training long-horizon, camera-controllable video world models.

The full technical report of AlayaWorld, a video world model that generates interactive, explorable, and continuously evolving virtual worlds from text, an image, or video.

Hand2Bot, an RGB-D video dataset with rich human context (body posture, facial expressions) collected for human-to-robot handover, plus RGB-D video generation to narrow the sim-to-real gap.

SAGE turns unlabeled Internet videos into training data for 3D geometric foundation models, combining sparse SfM anchors with dense consistency from 3D Gaussian rendering.

Monocular 3D detectors work best on images from the camera they were trained on. Decoupling the intrinsic parameters, by mapping images or features to an aligned focal length, lets one lightweight detector work across cameras.

For sparse-view inputs, an SDF supplies coarse geometry to improve 3DGS rendering, and images rendered by 3DGS in turn refine the SDF's fine surface detail.

Better instance-depth estimation for far-away objects (over 30m), which are small and rarely annotated, in monocular 3D detection.

Studies how driving knowledge in a vision-language model relates to closed-loop driving, and injects that knowledge into an end-to-end autonomous driving system.

A radar adaptive detection architecture that first classifies the interference covariance structure and then picks the matching adaptive detector.

Website template from Jon Barron.