Yu Zhang

Currently I work as a staff research scientist with SenseTime Research.

At SenseTime I've worked on SensePhoto, the state-of-the-art mobile photography solution dilivered to major smartphone OEMs. I earned my PhD at Beihang University, advised by Prof. Qinping Zhao and Prof. Bin Zhou. I also work as a visiting researcher at CVTEAM led by Prof. Jia Li. I did postdoc from 2019 to 2021, advised by Prof. Xiaogang Wang and Yebin Liu.

Email  /  CV  /  Bio  /  Google Scholar

I study computer vision, event-based vision, machine learning and optimization. My research lies much in image and video processing with learning and optimization methods.

(^†interns/students   *corresponding author)

By encoding the VFI prior into a few unfolded, learned gradient descent steps under the Bayesian regularization framework, our new VFI model achieves state-of-the-art results with only half the parameters of existing models, while showing better generalizability.

Human part segmentation can be conducted without dense pixel-level annotations by evolving a coarse part class map with image boundary cues, constrained by pose and object-level annotations.

Using an event camera allows you to train video interpolation models without the need of high frame-rate videos.

A fully trainable white-box event camera simulator with divide-and-conquer domain adaptation that automatically calibrate its parameters towards target domain.

Cross-domain pixel-level correspondences can be learned in weakly supervised manner for object detector adaptation.

Unpaired image translation from low light to day light can be achieved by the HDR of event streams captured by an event camera.

A dataset for aerial saliency detection and how to adapt existing saliency models to this task.

Simply learning feature super-resolution and knowledge distillation in multi-task way produces an accurate face detector capable of processing 763 faces/s on mobile phone.

An end-to-end learning pipeline that restores a motion blurred image to a video sequence using event camera.

Useful context information can be mined by optimizing the processing order of parts in semantic part parsing.

Semantic part parsing benefits from jointly processing multiple classes by attending to part boundaries and class discrimination.

Modeling the transitions across object boundaries helps salient object segmentation.

A new dataset for omnidirectional image stitching and novel metrics for assessing the quality of related algorithms.

Multiscale adversarial training on feature correlations defines unsupervised structural preservation loss for novel view synthesize.

By improving reference frame selection and 6D pose prediction, we reconstruct dynamic objects in real-time while handling large motion.

Modeling object part relations with simple priors enables accurate object localization in videos with weak supervision.

Exemplar detectors are explored to learn instance-specific saliency patterns and a large saliency detection dataset is proposed.

Extended version of CVPR 2015 with improved network flow solver and object shape prior.

Searching for the posed images with appearance similar to the input image can uniquely determine its pose with fast inference.

Shapes of local image patches lie in low-dimensional manifold, which is a consistency regularizer for image co-segmentation.

By incorporating global layout consistency modelled with maximum weighted clique, previous detection rate of cuboid proposals in RGBD images is doubled.

Weak object detectors can generate strong video object segmentation results via joint inference with a quadratic network flow model.

We present a fast approach for semantic segmentation by propagating labels along geodesic paths in feature space.

Yu Shi, Undergraduate student from Peking University, internship 2021-2022. Now a graduate student at UCLA.

Zhe Jiang, Master student from Sichuan University, intership 2018-2020. CVPR 2020 and ECCV 2020. Now a PhD student at The Hong Kong Polytechnic University.

Song Zhang, Master student from Beihang University, intership 2018-2020. ECCV 2020. Now a researcher at SenseTime.

Luwei Hou, Master student from Beihang University, intership 2020-2021. CVPR 2021. Now a researcher at SenseTime.