林佳縈 Chia Ying Lin
Computer Vision Reseacher
Computer Vision Lab, NTHU
With 4-year solid training in computer science fundamentals, 2-year independent research experience in computer vision, and diverse hands-on development experiences, I always pursue excellence, crave new challenges, and am open to all possibilities.
https://github.com/lykasbongbongbong
Skills
Languages
- Python (familiar)
- C++, C
- JAVA
Deep Learning Frameworks
- PyTorch (familiar)
- Tensorflow
CV & DL Libraries
- OpenCV
- Keras
Development Tools
- Git/Github
- RESTFul APIs
Web Development
- HTML, CSS, JS
- Python Flask
- PHP, Laravel
- MySQL, NoSQL(mongoDB)
- AWS
Others
- English TOEFL iBT 97
- English TOEIC 935
- German B2
Education
Master's Studies, Computer Vision Lab, NTHU
Dept. Information Systems and Application, Sep 2020~Now
- Anomaly detection and segmentation for smart manufacturing as primary research target
Bachelor Degree, WASN Lab, NCU
Dept. Computer Science, Sep 2016 ~ June 2020
Exchange Student, Hochschule München, Germany
Dept. Informatik (Computer Science), Sep 2019 ~ Feb 2020
Experiences
August 2021
Attendee,
Machine Learning Summer School, NTU
August 2021
Backend Developer (Anomaly Detection Demo Website),
CVLab, FUTEX 2021
- Use Python Flask, MySQL for RestFul APIs development
- Optimized backend system with Python threading to support sudden massive traffic
June 2018 ~ July 2019, 1y1m
Full-Stack Web Developer (NCU Internship Web)
Career Center NCU
- Reconstruct website with Laravel Framework
- Optimize backend and database to handle larger amounts of access
- Over 30% of users received intern opportunities via this site
- Website Link: https://ncuinternship.careercenter.ncu.edu.tw/
Master Thesis:
SABDN: Self-Attention Based Deviation Network for few-shot anomaly detection and segmentation (Under Review ECCV 2022)
Contributions
- Combine self-attention mechanism with feature extraction CNN network and anomaly synthesis mechanism for anomaly scoring to achieve outstanding anomaly detection accuracy under few-shot setting
- Reach SOTA performance on benchmark dataset MVTecAD dataset and BTAD dataset with over 90 percent reduction in training data requirements
Side Projects
GlueGAN: a generative model based on SuperGlue structure
- Extend MagicLeap’s SuperGlue end-to-end GNN concept for object synthesis to solve synthetic object image generation problem
- 10% accuracy improvement compared to initial GAN backbone
Let's play GAN with flows and friends!
- Generate synthetic object images with multi-label conditions with conditional GAN manner
- Human faces generation by conditional normalizing flow
2048: by Temporal Difference Learning Approach (RL)
- Construct TD-learning algorithm and design own n-tuple network to solve 2048 game
- Reach 98% 2048-tile win rate in 1000 games
The LunarLander: under Deep Q-Network and DDPG manner (RL)
- Implement DQN and DDPG network to solve LunarLander game problem
- Thoroughly understand Deep Q-learning, actor-critic mechanism
Real-Time car detection and counting system
- Retrain YOLOv3 with our own hatchback dataset
- Reach average accuracy: 99% on both day-light and evening scenarios
