Yansong Tang

I am a tenure-track Assistant Professor of Tsinghua-Berkeley Shenzhen Institute, Tsinghua University, where I direct the IVG@SZ (Intelligent Vision Group at Shenzhen, the sister group of the IVG at Beijing). Before that, I was a postdoctoral researcher at the Department of Engineering Science of the University of Oxford, working with Prof. Philip H. S. Torr and Prof. Victor Prisacariu. My research interests lie in computer vision. Currently, I am working in the fields of video analytics, vision-language understanding and 3D reconstruction.

I received my Ph.D degree with honour at Tsinghua University, advised by Prof. Jie Zhou and Prof. Jiwen Lu, and B.S. degree in Automation from Tsinghua University. I have also spent time at Visual Computing Group of Microsoft Research Asia (MSRA), and Prof. Song-Chun Zhu’s VCLA lab of University of California, Los Angeles (UCLA).

I am looking for self-motivated Master/PhD/Postdoc. If you have top grades or coding skill, and are highly creative and interested in joining my group, please do not hesitate to send me your CV and transcripts of grades.

Email  /  Google Scholar  /  GitHub

(*Equal Contribution, #Corresponding Author)

We present an end-to-end hierarchical Transformer-based network for referring segmentation.

We present Bi-level Neural Volume Fusion, which leverages recent advances in neural implicit representations and neural rendering for dense 3D reconstruction. In order to incrementally integrate new depth maps into a global neural implicit representation, we propose a novel bi-level fusion strategy that considers both efficiency and reconstruction quality by design.

DenseCLIP is a new framework for dense prediction by implicitly and explicitly leveraging the pre-trained knowledge from CLIP.

We devise a temporal-spatial Cubism strategy, which guides the network to be aware of the permutation of the segments in the temporal domain and the body parts in the spatial domain separately, thus improves the generalization ability of the model for cross-dataset action recognition.

Journal version of the COIN dataset.

We observe a collapse phenomenon when directly applying fully convolutional cycle-consistency method for video correspondence learning, study the underline reason behind it, and propose a spatial transformation approach to address this issue.

We present an end-to-end hierarchical interaction network for video object segmentation from referring expressions, which leverages the feature pyramid produced by the visual encoder to generate multiple levels of multi-modal features.

We propose an uncertainty-aware score distribution learning method and extend it to a multi-path model for action quality assessment.

We propose a graph interaction networks (GINs) model for transferring relation knowledge across two graphs two different scenarios for video analysis, including a new proposed setting for unsupervised skeleton-based action recognition across different datasets, and supervised group activity recognition with multi-modal inputs.

We extend of our Semantics-Preserving Attention model with graph convolutional module for group activity recognition.

COIN is one of the largest and most comprehensive instructional video analysis datasets with rich annotations.

We propose a multi-stream deep neural networks and THU-READ dataset for RGB-D egocentric action recognition.

We present a simple yet effective semantics-preserving attention module for group activity recognition.

We propose a simple yet effective method to select key frames for skeleton-based action recognition using the REINFORCE algorithm.

Website Template