KeJia Yin

ICON above was generated by DCGAN by myself :)

2022 MScAC Student at University of Toronto | First Author Publication at CVPR 2024 | Actively seeking Machine Learning Researcher / Applied Researcher / Machine Learning Engineer

Welcome to my personal homepage

•

RESEARCH EXPERIENCE

Publication:
  • Kejia Yin, Varshanth Rao, Ruowei Jiang, Xudong Liu, Parham Aarabi, David B. Lindell, “SCE-MAE: Selective Correspondence Enhancement with Masked Autoencoder for Self-Supervised Landmark Estimation”, in IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2024.
    •

    SCE-MAE: Selective Correspondence Enhancement with Masked Autoencoder for Self-Supervised Landmark Estimation | May 2023 – Nov 2023

    Advisor: David B. Lindell, Assistant Professor at Department of Computer Science, University of Toronto

  • Proposed a novel method addressing self-supervised facial landmark estimation problem by utilizing Masked Autoencoder and selective correspondence enhancement, and implemented this method with Pytorch.
  • Reduced the error by 20~44% and 9~15% on landmark matching and landmark detection respectively compared with the previous state-of-the-art methods.
  • • Composed a conference paper which is accepted to CVPR 2024 as the first author.
    •

    Research on Domain Generalization Image Classification Based on Gaussian Kernel | Dec 2021 – May 2022

    Advisor: Shuang Li, Assistant Professor at the School of Computer Science, Beijing Institute of Technology

  • Proposed a novel method addressing domain generalization image classification problem by utilizing gaussian kernel to extract the high-frequency information from the image, and implemented this method with Pytorch.
  • Achieved 6.2% and 4.52% mean classification accuracy improvement on Digits-DG and PACS dataset respectively compared with the baseline method, which were competitive results compared to state-of-the-art methods.
  • Wrote a thesis by myself and defended it with five professors from the Department of Computer Science.
    •

    Balancing real-world inverted pendulums via virtual training with RL | Jul 2021- Aug 2021

    Advisor: Hien Tran, Professor and Associate Head at Department of Mathematics, North Carolina State University

  • Implemented Policy Gradient, Actor Critic, and Proximal Policy Optimization with Pytorch and successfully balanced the single inverted pendulum in a modified gym environment which provides a more realistic simulation of physical laws.
  • Successfully balanced the double inverted pendulum by using Actor Critic in a modified gym environment.
  • Successfully applied our trained model in the gym to balance a real single inverted pendulum in the lab.
  • Trying to directly apply our trained model in the gym to balance a real double inverted pendulum in the lab.
    •

    Long-tailed image classification | Feb 2021- Mar 2021

    Advisor: Shuang Li, Assistant Professor at School of Computer Science, Beijing Institute of Technology

  • Explored novel approaches to addressing long-tailed image classification problems by augmenting long-tailed datasets via synthesizing new data according to closest neighbor classes.
  • Implemented a novel method and evaluated it on artificial long-tailed datasets CIFAR-10 which had competitive results with state-of-the-art methods
    •

    Theoretical understanding and image generation with GAN | Nov 2020 - Dec 2020

    Advisor: Shuang Li, Assistant Professor at School of Computer Science, Beijing Institute of Technology

  • Implemented existing works on generative adversarial networks for images with Pytorch including DCGAN, WGAN, Cycle GAN.
  • Wrote a report to analyze the causes of instability of training GAN from the aspect of loss function based on formula derivation and explain the advantages of the loss function used in WGAN.
    •