Hamilton Chang's Blog
0%

About Me

Chang Huai-Yuan

Email: hamiltonchangwork”at”gmail.com / hamiltonchang”at”pku.edu.cn

Education

  • Peking University, M.E. in Software Engineering, Sept. 2017 - Aug. 2020

  • Chung Yuan Christian University, B.S. in Computer Science and Information Engineering, Sept. 2013 - June 2017

Experience

  • Infortrend Technology, Taiwan, New-Taipei

    AI Research and Develop Engineer, May.2021-now

    • Led project AI Service API and project Resource Space
      • Utilized Trition Inference Server with ONNX, TensorRT, and Python backend
      • Developed Face Recognition API:
        • Regarded MTCNN as the basic detection model, providing FaceNet and ArcFace as recognition options
        • Used shared memory to solve the issue of multi-processes shared face database issues and improved performance by 3 times
        • Utilized quantization and TensorRT to get 5 times performance
      • Utilized Multi-Stage, deleted redundant dependencies, etc. to reduce half docker image size
      • Reduced 3 times inference time on general CPU platforms, using TensorFlow Lite with XNNPack
    • Developed an auto-tiered model to improve IO performance by 60 % on Infortrend storage systems
      • Developed a classification model to predict if files are accessed in the next period
      • Accuracy of the classification model only gets 80% at most if the next period is a day
      • Developed a regression model to predict the heat of files and improve at least 60% IO performance of the storage system

  • Patere Technologies, Inc., Taipei, Taiwan

    Computer Vision Engineer, Nov. 2020 - Jan. 2021

    • Confidence and speech fluency detection
      • Surveyed how to detect the speech fluency and confidence of the candidate with speech detection
      • Designed and implemented the rule-based project with basic speech features. Researched more complex and useful features which served for ML/DL, such as MFCC, FBank features
    • Focus detection
      • Designed and Implemented the algorithm to check the candidate is focusing or not
      • Accuracy, precision, and recall are all about 80% on the testing dataset

  • Lenovo Group Ltd, Lenovo Research, Beijing, China

    Computer Vision and AI Algorithm Intern, Aug. 2019 - Nov. 2019

    • Developed automatic training system for commodity detection based on RetinaNet for unmanned stores, and found the reason resulted in low detect success rate by analyzing the testing result
    • Increased the success rate of commodity detection about 20% through collecting and generating more complicated and suitable data

  • Zero Zero Robotics, Beijing, China

    Computer Vision and AI Algorithm Intern, Nov. 2018 - Aug. 2019

    • Involved in Hover 2 drone development
      • Researched and tested correlate filter and Template Matching Algorithms, and analyzed them
      • Implemented, maintained and optimized long-term object tracking function
      • Improved by 15% tracking success rate and reduced by 5% CPU utility through adding constraints and modifying object re-identification strategy

Competition

  • IBM Watson Build Competition: Clothes Master

    Team Leader, May 2018 - June 2018

    • 2nd place in China
    • Implemented a personal dress recommendation web which based on IBM Watson chatbot
    • Designed software architecture and main functions, and Implemented front-end(Javascript) and back-end(Node.js)

Project

  • Design and Implementation of Real-time Long-term Single Person Tracking System

    Individual Work, Nov. 2019 - April. 2020

    • Divided the tracker into three states: target selection, track and retrieval, to facilitate the design and implementation
    • Designed and Developed several strategies to achieve real-time and stable long-term person tracking, such as skip-frame, conflict detection, tracker constraint scheme, and exclud e target range
    • Tracking successful rate is 79.02%. Recall is 77.54%. FPS is 12.87 on Intel i7-6700k CPU

  • 2D Anime Charactors Recognition

    Team Leader, May 2018 - June 2018

    • To recognize the anime character by the same character but different styles
    • Utilized the trained LBPcascade to detect anime faces and recognized ones with the fine-tuned Inception-v3 model.
    • TOP1 result is 73.25%

  • Scheme Interpreter Implemented in C++

    Individual Work, Mar. 2017 - May 2017

    • Designed Scanner and Parser, and Binding and Error Message by analyzing Scheme code structure
    • Transformed Scheme code into designed logical tree structure

  • Face Recognition Access Control System Based on Raspberry Pi

    Team Leader,  Feb. 2016 - Nov. 2016

    • Learned and used the technologies and pre-processing of face recognition, and balanced the security and practicality of the system
    • Built the software system containing face key(database) and face recognition, and ported it to Raspberry Pi

Skills

  • Programming Language: Python, C/C++, Java, R
  • Other: Git, GDB, Docker, pdb, Kubernetes