Career Profile

Highly motivated, self-driven Computer Vision / Backend practitioner with 1.5+ yrs research experience in Deep Learning and Computer vision. Participated in multiple related projects of deep learning, computer vision, data mining and back end.

Related Experience

Research Assistant

12/2017 - current, BALTIMORE, US
Johns Hopkins University

Work Description:

Participated in Deep Intermodal Video Analytics (DIVA) project at Johns Hopkins University with following responsibilites:

1. Designed, implemented and fine-tuned vehicle semantic part detection / pose estimation neural network and data pipeline.

2. Synthesized car images using graphical rendering engine (i.e. Unreal Engine) with spare key point annotation, plus domain randomization for bridging the gap between synthetic data and real (DIVA) data.

3. Provide cues of temporal shape dynamics and functional relationship between cars and humans for action recognition.

4. Perform object (car and human) tracking with feature embedding algorithms and deep nets (e.g. triplet loss network) using hard data mining strategy and particle filter (SMC - Sequential Monte Carlo) / Kalman filtering graphical model.

Link: here

Research Assistant

07/2018 - 09/2018, BALTIMORE, US
Johns Hopkins University

Work Description:

Participated in Measuring Patient Mobility in the ICU (ICU) project at Johns Hopkins University with following responsibilites:

1. Design hierarchical activity filtering system and pose/appearance based feature extraction + temporal spatial classifier for activity classification in ICU.

2. Train activity classifier with few-shot and semi-supervised based methods to overcome the scarcity in training data.

Link: here

Research & Course Projects

(Research Project) Robust Vehicle Semantic Part Parsing & Pose Estimation (in DIVA project) - A robust, occlusion aware vehicle semantic part parsing framework. It is a subtask in Deep Intermodal Video Analytics project which regresses the 2D and 3D semantic part locations jointly. The combination of novel network architecture design, synthetic data pipeline with augmentation and loss function design has improved the model performance significantly especially under occlusion/truncation situations (which happens quite often in real scenarios). Please refer to: Diva Project Highlight, Code and My Presentation Slides.
(Research Project) Triplet Network for Human Re-identification (in DIVA project) - This project aims at dealing with human tracking failure especially for extended frames span in DIVA project. By training a triplet neural network to encode pedestrians appearance feature and re-identify pedestrians, improved ROC result is achived. More info please refer to Code and Presentation Slides.
(Course Project) Food_Mate - This is a course project of JHU Object Oriented Software Engineering class, where an Android app with developed for satifying people's need for dating and having delicious meal. It includes frond end, back end, CI, unit testing etc. and the full life cycle of app development(requirements; design; implement; test; deploy) has been covered. The prototype of this app can be found: here. Here is presentation slides: Food Mate Presentation.
(Course Project) Classification of Disadoption of Water Efficient Technologies in Rural Costa Rica - This is a course project of JHU Data Mining class, cooperated with JHU Environment Engineering Department, Carey Business School and Maryland University. We collected and analzed data for Puerto Rico customer data and trained our machine learning model to classify customer behaviors. Please refer to: Code and Presentation.