Zheng-yuan Yang

I am a Ph.D. candidate in Computer Science at University of Rochester, Rochester, NY, advised by Prof. Jiebo Luo. I recieved my B.E. degree from University of Science and Technology of China in 2016.

This site is still under constructing, more projects will be posted.

Currently I'm looking for summer interns for 2019 on Computer Vision (Image/Video Analysis).

Email  /  CV  /  LinkedIn

Publications

  • Zhengyuan Yang, Yuncheng Li, Jianchao Yang, Wendi Cross, Jiebo Luo, “Pose-based Body Language Recognition for Emotion Interpretation,” Submitted to AAAI 2018.
  • Zhengyuan Yang, Wendi Cross, Jiebo Luo, “Personalized pose estimation for body language understanding,” International Conference on Image Processing (ICIP), Beijing, China, September 2017. (oral presentation)
  • Research

    I'm interested in computer vision, machine learning, data mining, image processing, and multimedia. Much of my research is about understanding content from images or image sequences, e.g. video action recognition and pose estimation. My works also related with data mining.

  • Click for zooming up.
  • Pose-based Body Language Recognition for Emotion Interpretation
    Zhengyuan Yang, Yuncheng Li, Jianchao Yang, Wendi Cross, Jiebo Luo
    Submitted to AAAI, 2018
    PDF to appear

    Inspired by human ability to infer emotion from body languages, we propose a pose-based body language recognition framework for emotion interpretation from videos. The proposed system in the first stage generates sequences of body language predictions on untrimmed videos with multi-modality joints information supplied by a reliable pose estimator called OpenPose. In the second stage, the predicted sequences are fed into a temporal network for emotion interpretation.

    Personalized Pose Estimation for Body Language Understanding
    Zhengyuan Yang, Wendi Cross, Jiebo Luo
    ICIP, 2017
    PDF / Slides / Structure / Results

    A novel model is proposed for analyzing psychiatry conversations with pose estimation methods. Various modifications are designed specific for the target dataset, which include less motion while require higher accuracy and reliability.

    Contributions include:

  • We propose a fully convolutional spatial fusion architecture, in contrast to graphical models, to encode the spatial relations between the joints.

  • We introduce a personalized annotation step to boost the reliability of the joint predictions. The percentage of implausible annotations is iteratively reduced.

  • We establish a psychiatric conversation dataset for computer vision based subtle body language extraction and mental disorder detection.
  • Steering Angle Control with End-to-end Neural Networks
    Zhengyuan Yang, Jerry Yu, Junjie Cai, Jiebo Luo
    Slides / Demo

    Taking an image or image sequences as input, predict steering angle end-to-end with neural networks.

  • Click for zooming up.
  • Multi-view Visual Recognition with Knowledge Distillation
    Zhengyuan Yang, Yuncheng li, Jiebo Luo
    PDF to appear

    Propose to use knowledge distillation to fuse the information from the multiple views of data and ground-truth labels. A supervised multi-view learning framework is proposed to train the visual classifiers.

  • Click for zooming up.
  • Oral or Poster? Publication Influence Analysis
    Zhengyuan Yang, Jiebo Luo
    PDF

    We analyze the influence of presentation type on academic publications in the computer vision field:

  • Collecting information and citation network for more than 10,000 publications.

  • Analyzing the influence of presentation type (Oral, Poster) in detail.

  • Proposing a ranking based representation for measuring publication influence.
  • Click for zooming up.
  • Action Based Marine Animal Recognition
    Zhengyuan Yang, Mohammed Bennamoun
    Poster

    We propose a method to classify marine species in videos based on the features of their movements:

  • Combining ConvNet and hand-craft features for spatial feature extraction.

  • Introducing a nonlinear regression to extract temporal information as the feature vector for each video.

  • Achieving competitive performance on public action recognition datasets Hollywood2, HMDB51, and high accuracy on real marine animal data.
  • Internship

    SAIC Innovation Center, San Jose
    Jun - Aug 2017. Advisor: Jerry Yu.

    Teaching

    TA CS172 - Fall 2017
    Datastructures and Algorithms

    TA CS242 - Spring 2017
    Intro to Artificial Intelligence


    © 2017 Zhengyuan Yang. All rights reserved.
    Template borrowed from Jon Barron. Thanks!