JING-AN TZENG

JING-AN TZENG

Video Processing Algorithm Engineer

Zoom Video Communication Inc.

Interests

  • Computer Vision
  • Artificial Intelligence
  • Image Processing
  • Machine Learning
  • Robotics

Education

  • M.S. in Electrical and Computer Engineering, Sept 2019 - Jan 2021

    University of Michigan

  • B.S. in Power Mechanical Engineering, Sept 2014 - Jun 2018

    National Tsing Hua University

Experience

 
 
 
 
 

Zoom

Software Engineer, Video Processing Algorithm Engineer

Jul 2021 – Present San Jose, CA

  • Conducted comprehensive research, design, training, implementation, and evaluation of a lightweighted virtual background model (VB) specifically optimized for low-end machines.
  • Achieved 2x speed improvement compared to our standard VB model and reduced the size to less than 100KB.
  • Developed an asynchronous face recognition pipeline utilizing three distinct deep learning models: object detector, person re-identification, and face recognition.
  • Prototyped face recognition pipeline for Zoom Room on Windows and MacOS platform in C++.
  • Facilitated communication and collaboration among teams and across different companies.
 
 
 
 
 

Advanced Computer Vision – Improvements on Object Detection (Faster R-CNN)

Software Team Leader

Sep 2020 – Dec 2020 Ann Arbor, MI

  • Trained and evaluated the Faster R-CNN on PASCAL VOC 2007 with PyTorch.
  • Decoupled the classification head and localization head and improved the mAP by 1.
  • Introduced efficient channel attention and cross stage network to ResNet-50 backbone and increased 5 mAP..
 
 
 
 
 

Individual Study – Benchmark for Video Inpainting

Team Member

Sep 2020 – Dec 2020 Ann Arbor, MI

  • Designed algorithms to classify the degree of attributes — camera motion, foreground motion and displacement — which affect the video inpainting model performance on DAVIS dataset.
  • Evaluated the performance of the algorithm by Precision – Recall curve and ROC curve.
 
 
 
 
 

ASML

Droplet Generation Control and Automation System Intern

May 2020 – Aug 2020 San Diego, CA

  • Designed an object detection pipeline from scratch by OpenCV, Tkinter and Scikit-learn in Python to detect the tin droplets and satellites for preprocessing, labeling, feature extraction and classification.
  • Detected the interest objects with maximally stable extremal regions (MSER) and eliminated the overlapped bounding boxes with Non-Maximum Suppression (NMS).
  • Evaluated the classifiers’ performance with k-fold cross validation, learning curve and validation curve.
 
 
 
 
 

Mobile Robotics - Visual Inertia Navigation (VI-SLAM)

Team Leader

Mar 2020 – Apr 2020 Ann Arbor, MI

  • Improved a Muti-State Constraint Kalman filter-based visual inertial navigation framework (Openvins) with learning- based interest point extractor – SuperPoint in Pytorch by using C++.
  • Evaluated the performance on the EuRoC MAV dataset with ROS and ameliorated the performance for every tasks.
 
 
 
 
 

Computer Vision - Depth Completion

Team Member

Mar 2020 – Apr 2020 Ann Arbor, MI

  • Completed dense depth data from a color image and sparse LiDAR data in KITTI depth completion benchmark.
  • Developed two-pathway learning architecture with U-Net like low-resolution feature extractor and utilized attention mechanism to propose the final prediction in Pytorch.
 
 
 
 
 

Self-Driving Car - Object Detection (YOLOv3)

Perception Team Leader

Nov 2019 – Dec 2019 Ann Arbor, MI

  • Placed 2nd overall in class and implemented and trained YOLOv3 on a given dataset in Python with Keras.
  • Developed a layer to discriminate the distance of the objects with point cloud.
 
 
 
 
 

Eurobot 2018 Contest - Autonomous Robot

Software Team Leader

Sep 2017 – Jun 2018 Taiwan/France

  • Placed 24th overall in world counted.
  • Prototyped positioning system model with MATLAB for integration test, utilized microcontrollers and ultrawideband (UWB) chips to locate the robots accurately and practiced the whole system with C++ in Linux.
  • Implemented Kalman filter and trilateration algorithm to enhance measuring accuracy and stability, increasing 30% accuracy which is within 5 cm radius with high repeatability.
 
 
 
 
 

Mobile Robot

Control Team Leader

Sep 2017 – Jan 2018 Hsinchu, Taiwan

  • Developed a discrete-time PID controller and implemented it on Altera DE0-Nano by programming in Verilog.
  • Handled input from rotary encoders and output with PWM signals to control the rotational speed
 
 
 
 
 

Wafer Gripper

Software Team Member

Sep 2017 – Jan 2018 Hsinchu, Taiwan

  • Designed a G-Code Interpreter to read the NC (numerical control) code and execute them on 3-axis wafer gripper.
  • Built a human - computer interface and stimulated the wafer gripper’s motion by using Windows Forms.
  • Implemented trapezoidal velocity profile and interpolation to control the stepper motors smoothly.
 
 
 
 
 

Powered Exoskeleton for Motion Recording

Sensing and Communication Team Leader

May 2017 – Dec 2017 Hsinchu, Taiwan

  • Placed 2nd and “Most Popular” awards from 50 teams in Senior Capstone Project Competition.
  • Developed a sensing exoskeleton suit consisting of 4 microcontrollers, 6 encoders and 4 inertial sensors for detecting and recording the user’s movements, i.e. hand gestures.
  • Used SPI, I2C, UART for fetching the data to the microcontrollers and fused the gyroscope and accelerometer data by compensation filter.
  • Built a Bluetooth communication system with cyclic redundancy check (CRCs) to improve its consistency.

Projects

*
All Robotics Computer Vision Machine Learning

Improvements on Object Detection (Faster R-CNN)

Modified the backbone, detection head and RPN to imporve the performace of Faster R-CNN

Video Inpainting Benchmark

Divide the videos’ motion into binary labels and evaluated those videos’ difficulties for the video inpainting task.

Depth Completion

Developed a learning structure to estimate a dense depth image from sparse depth measurements.

Visual Inertia Navigation

Navigated the drone with Openvins and the learning based feature extractor – Superpoint

Eurobot 2018

Biult two robots and solved the tasks in Eurobot which is an international robotics contest in Europe annually.

Exoskeleton

Seeked a new way to deal with the shoulder rehabilitation problem

Courseworks

Advanced Computer Vision

University of Michigan Sep 2020 – Dec 2020
  • Convolutional Neural Network (CNN)
  • Object Detection
  • Image Processing

    • Mobile Robotics

    University of Michigan Jan 2020 – Apr 2020
  • Simultaneous localization and mapping (SLAM)
  • Kalman Filtering , Particle Filtering
  • Point Cloud Registeration

    • Foudamental of Computer Vision

    University of Michigan Jan 2020 – Apr 2020
  • Convolutional Neural Network (CNN)
  • Object Detection
  • Image Processing

    • Embedded Control

    University of Michigan Jan 2020 – Apr 2020
  • Digital Control System
  • Embedded System

    • Computational Data Science and Machine Learning

    University of Michigan Jan 2020 – Apr 2020
  • Singular Value Decomposition (SVD)
  • Least Squares Regression

    • Self-Driving Cars: Perception and Control

    University of Michigan Jan 2020 – Apr 2020
  • MPC Control
  • Deep Learning

    • Control System

    National Tsing Hua University Jan 2018 – Apr 2018
  • PID Control System
  • Linear Control System

    • Contact