1. Work Project

Whole lifecycle Real Time Pose Estimation System, 06/2020 - 06/2023

• Ensure accurate and refined pose estimation by seamlessly integrating Human Feedback into the pose estimation model pipeline. Optimize workflow and maximize productivity through development of a highly efficient full-cycle deep learning system pipeline with MLops. Expand pose dataset to a massive TB level by incorporating fish-eye images and annotations. Initiate productive discussion and delivered tactful negative feedback to manager during half year performance review to prompt subsequent changes in working methods, ultimately resulting in enhanced team performance.
• Enhanced model inference to an impressive 30 FPS by efficiently loading and running multiple Models on a single GPU.
• Explored cutting-edge techniques and advancements in field by conducting in-depth research on state-of-the-art pose
  estimation models.
• Reduced Model training and delivery time significantly by leveraging cost-effective GPUs without compromising on
  performance and accuracy.
• Fostered transparent communication with clients and internal team, leading to a two-week deadline extension and the
  subsequent on-time completion of the project following the early delivery of a functional demo version.
• Achieved an impressive 90% person tracking accuracy through successful implementation and deployment of real-time 2D Pose Estimation, 3D Pose Tracking, and person re-identification models for a checkout-free retail system.
• Resolved disagreement with Team Manager by executing experiments and engaging in discussions to identify a hybrid
  approach, effectively reducing model training and deployment cost for pose estimation and tracking.
• Related Patents:
  • US-11551079-B2 | US-11270260-B2

2. Computer Vision Project

Face Anti-Spoofing Detection on Mobile Device, 03/2019 - 04/2019

• Develop real-time Face Anti-Spoofing Algorithm and Mobile System.
• Research and Develop Face Recognition and Action Recognition Algorithm on Mobile System.

Human Pose Estimation and Tracking on Mobile Device, 04/2019 - 06/2019

• Related Publication: [Bounding Box Embedding for Single Shot Person Instance Segmentation](https://arxiv.org/pdf/1807.07674v1.pdf)
• Optimize real-time Deep Learning based Pose Estimation, Tracking and Instance Segmentation algorithm on Mobile System with Flow-based Pose Tracking Algorithm.

Gaze Estimation and Tracking, 05/2019 - 08/2019

• Develop real-time Deep Learning based Gaze Estimation System with 10ms latency and less than 5 mean degree error on MPIIGaze, UnityEyes datasets.
• Design and Collect OctiGaze dataset that contains 256,876 images and gaze vectors we collected from 101 participants in both RGB and IR camera.

ECCV 2018 Workshop - WIDER Pedestrian Detection Challenge, 05/2018 – 07/2018

• Achievement: 1st in Development Phase, 12nd in Final Phase.
• What I Do: 1) Build deep learning models for Pedestrian Detection tasks with Detectrion and Tensorflow Framework. 2) Implement Cascade RCNN based on Faster RCNN. 3) Implement easy-to-use tools for image and result analysis.
• Key Words: Object Detection, Pedestrian Detection, Faster RCNN, Cascade RCNN, YOLO, Deep Learning.

3. Data Mining Project in Kaggle

Competition Contributor 01/2016 - NOW

WSDM 2018 WORKSHOP - KKBOX'S CHURN PREDICTION CHALLENGE, 9/18/2017-12/19/2017

• Achievement: TOP 8%
• What I Do: 1) Build multiple models to predict whether a user will churn after their membership expires via
  XGBoost, LightGBM, AutoEncoder. 2) Propose a method to do large file feature engineering based on
  “Equivalence Substructure” idea.
• Key Words: Feature Engineering, Autoencoder, Large Data Processing, Model Stacking

INSTACART MARKET BASKET ANALYSIS, 07/2017- 09/2017

• Achievement: Bronze Medal, TOP 10%.
• What I Do: 1) Build models to predict which previously purchased products will be in a user’s next order. 2)
  Use F1 Optimization for post processing.
• Key Words: Feature Engineering, XGBoost, LightGBM, Model Stacking, F1 optimization.

MERCEDES-BENZ GREENER MANUFACTURING, 07/2017- 07/2017

• Achievement: TOP 13%
• What I Do:  1) Build model to predict the time it takes to pass testing.
• Key Words: Word Embedding, GradientBoostingRegressor, Lasso, LightGBM.

SBERBANK RUSSIAN HOUSING MARKET, 06/2017- 08/2017

• Achievement: TOP 15%.
• What I Do: 1) Build model to predict realty house prices
• Key Words: Feature Engineering, XGBoost, LightGBM, Model Stacking.

QUORA QUESTION PAIRS, 09/2017- 12/2017

• What I Do: 1) Build a LSTM model to decide whether two questions have similar meaning.
• Key Words:  Word Embedding, Word2Vec, Representation-based and Interaction-based Model, LSTM-RNN.

4. Research on 3D Face Tracking & Recognition Technology based on Depth Data

Chief Researcher 05/2014- 04/2015

Publication:

• Research on the 3D face recognition based on multi-class classifier with depth and point cloud data 10.1109/IMCEC.2016.7867242

• Research of 3D face recognition algorithm based on deep learning stacked denoising autoencoder theory 10.1109/ICCSN.2016.7586606

Patents:

• A New Type of Face Deep Data Recognition Algorithm and Application in Intelligent Authority Distribution System, No.: 201510232302.6;
• One 3-D Face Recognition Algorithm based on Deep Learning SDAE and Application in the Finance Field, No.: 201510145327.2;

5. Research on Non-contact Medical Assistance Detection System of Elderly Parkinson's disease

Chief Researcher, 04/2015-04/2016

Patents:

• One Human Behavior Recognition Method Based on Manifold Learning, No.: 201510179381.9;
• Methods of judging fall behaviors and body balance for old people, No.: 201410633427.5;
• Gait Recognition and Quality Assessment algorithm for elderly Parkinson’s disease based on skeleton data, No.: 201610161393.3;