Yu Zheng

I am currently a Postdoctoral Research Fellow of Department of Automation at Tsinghua University, affiliated with Intelligent Vision Group (IVG), supervised by Prof. Jie Zhou.

I received the B.Eng. and Ph.D. degrees in Control Science and Engineering from Department of Automatjion, Tsinghua University, China, in 2019 and 2024, respectively, supervised by Prof. Jiwen Lu.

My recent research interests focus on computer vision safety and 3D vision.

Email  /  GitHub  /  Google Scholar

We enhance the 3D representational ability of Neural Radiance Field through artificial sight degradation and self-supervised recovery.

It is the first 3D Gaussian based 3D object detector, where we address the outlier issue via closure measurement.

We propose a graph-based 3D point cloud recognition scheme that weighs the contribution of local geometries from an entropy perspective.

We propose to decouple the 3D object detection task into ''what'' and ''where'' as an upstream scene prior .

It is the first generalizable semantic radiance field, where we tackles the undue multi-view semantic ambiguity by designing a cross-reprojection attention.

We propose to leverage the low-noise representational ability of low-resolution point cloud to regularize the high-noise high-resolution point cloud.

We propose to provide scene prior for 3D objects that share geometry features but are similar in color or texture.

We propose to fully utilize the synthetic 3D assets with physics constraint to serve for real-world 3D detection with scarce supervisions.

An optimizing method that tackles the loss-metric mismatch in oriented object detection, not limited to 3D scenarios.

We propose to learn the abundant geometrical and locational relations in 3D point cloud.

We construct by far the largest-scale (>10k) instructional video datasets with abundant annotations.

Website Template