Me
Tsai-Shien Chen
Graduate Researcher
at National Taiwan University
tschen@media.ee.ntu.edu.tw

Biography

Tsai-Shien Chen obtained his bachelor's degree in Electrical Engineering at National Taiwan University. Subsequently, he was a master student and graduate researcher in Electronics Engineering at National Taiwan University where he conducted research at Media IC and System Laboratory advised by Prof. Shao-Yi Chien. If you would like to learn more about him, please see his [CV] or contact him at tschen@media.ee.ntu.edu.tw

Research Interests

  • Machine Learning / Deep Learning
  • Computer Vision
  • Image Classification and Recognition
  • Vehicle / Person Re-Identification

Education

National Taiwan University (NTU)
  • Bachelor in Electrical Engineering, 2019
    • Overall Class Rank: 5th / 190
    • Overall GPA: 4.23 / 4.30

  • Master in Electronics Engineering, Ongoing

Timeline & Experiences

2019 - (Ongoing)
Graduate Researcher @ NTU
Master in Graduate Institute of Electronics Engineering
Media IC and Systen Lab
Primary focus:
- Computer Vision
- Image Classification and Recognition
- Vehicle/Peron Re-Identification
See publications.
Supervisor: Prof. Shao-Yi Chien
2019 Summer
Scientific Research Intern @ MediaTek
Explored a deep learning based algorithm for video encoding.
2017 Summer
Software Engineering Intern @ ITRI
Developed a software tool to simulate the force analysis.
2015 - 2019
Undergraduate Student @ NTU
Bachelor in Department of Electrical Engineering
See projects.
1996
He was Born

Publications

Orientation-aware Vehicle Re-identification with Semantics-guided Part Attention Network
In this paper, we proposed a network, named SPAN, to predict the spatial attention map for each vehicle view given only image-level label for training. Then we introduce a distance metric which places emphasis on the co-occurrence vehicle views when evaluating the feature distance between two images. The experiments showed our superiority in both the performance of re-identification and the quality of generated attention maps compared to the state-of-the-arts.
European Conference on Computer Vision (ECCV), 2020 [Oral]
Viewpoint-Aware Channel-Wise Attentive Network for Vehicle Re-Identification
In this work, we proposed an attention mechanism, named VCAM, which enables our framework channel-wisely reweighing the importance of each feature map according to the viewpoint of input vehicle image. By the aid of VCAM, we obtained promising results on the 2020 AI City Challenge. We also conducted the experiments to demonstrate the interpretability of how VCAM practically assists the learning.
Computer Vision and Pattern Recognition (CVPR) Workshops, 2020
Supervised Joint Domain Learning for Vehicle Re-Identification
To make our framework more efficient to train on multiple datasets simultaneously, in this paper, we proposed a network, named JDRN, to mitigate the domain gap due to misaligned feature distribution between different datasets. With our JDRN, we can disentangle domain-invariant information and encourage a shared feature space among domains.
Computer Vision and Pattern Recognition (CVPR) Workshops, 2019

Honors & Awards

International

Domestic

  • Valedictorian / Department of Electrical Engineering, National Taiwan University / 2019
  • 4 times of Presidential Awards / National Taiwan University / 2015-2019
  • 2nd Place / Deep Learning for Computer Vision: Final Project Contest / 2019
  • Ranked 4th (out of 200+) / Data Structure and Programming: Final Project Contest (hosted by Cadence Inc.) / 2018
  • Semifinal (Top 5%) / International Physics Olympiad Domestic Semifinal / 2014

Selected Projects

Vehicle Re-Identification & Traffic Anomaly Detection System (2018-2020)
Designed a system which can match vehicle images of same identity captured from different cameras and can also detect anomalies, such as lane violation, illegal U-turns and wrong-direction driving, etc. Got promising ranking in both 2019 and 2020 AI City Challenges and the papers were accepted for publication.
Worldwide Kaggle Competition: Human Protein Atlas Image Classification (2019)
The objective of this competition is to develop models capable of classifying mixed patterns of proteins in microscope images. To solve the problem of multi-label classification on 27 highly imbalanced classes, we proposed an algorithm with AdaBoost and ensemble technique to cope with imbalanced dataset and ranked 1st in class / 279th in the world.
Speech Recognition System (2019)
Built a complete speech processing system to classify the command type for the given speech clip (in wav file). The whole system mainly includes three steps: transformation from a raw speech signal into spectrogram, computation of 39-dim MFCC, and then classification of command type by a CNN model.
Speago: Voice Control Outfit Recommendation System (2017)
Implemented a smart closet which is controlled by an Android app and would automatically pick up the recommended outfit based on the weather, temperature and the voice command of the user.
Integrated Circuit Design: from Software to Hardware Development (2017)
In this project, we experienced the complete process of IC development, which mainly includes five steps: software design, RTL implementation, synthesis (to Gate-level), placement and routing, and finally taping out our custom chip. We also executed corresponding verifications for the intermediate and finally products after each step.
Rectifier (2017)
Designed a rectifier which can transfer the 110V AC input to 0~2.5V DC output and drive the mini fan with controllable rotation speed. In this project, we went through the ciruit design and printed circuit board (PCB) making and verification.
Top