Computer Vision Engineer  •  Academia Sinica

. Developed and improved many state-of-the-art deep learning models CNN, C3D, Siamese Network, Transformer, and YOLO series in Python3 and PyTorch.
. Top Achievement : NeighborTrack Che 22, the most accurate single-object tracking method in the world.
. Research scope : Computer Vision : Object detection/tracking, Person Re-Identification and Video Stabilization.

Projects:

Single object tracking 03 2021
I.I.S. Research, Framework:python/pytorch
• Designed a post-processing method NeighborTrack[CVPR2023] to introduce neighbor and temporal
information to alleviate the error tracking of single object tracking.
• Proved NeighborTrack is the state-of-the-art single-object tracking model as the accuracy on LaSOT is
72.2% AUC. NeighborTrack: Single Object Tracking by Bipartite Matching With Neighbor Tracklets and Its Applications to Sports [code]

Multiple object tracking 08 2019
I.I.S. Research, Framework:python/pytorch
• Used multi-scale features and non-local net in unknown class multiple object tracking to Improve base
method accuracy.
• Improved the base model by 1.2x Average Precision (33% to 40%) in MOT17 dataset.


Video based fall detection 04 2019
I.I.S. Research, Framework:python/tensorflow
• Implemented optical flow features and data augmentation to improve the accuracy of C3D-pelee deep
learning network in fall detection tasks.
• Increased the accuracy of the basic network, UCF101 dataset from 57.1 to 59.5, MCF dataset from 85.4 to
87.5.

Video person Re-ID 04 2018
I.I.S. Research, Team work, Framework:python/tensorflow on embedding system Jetson TX2
• Adapted the mobilenetV2 person ReID system to the embedded system Jetson TX2, which has only 7% of
the computing power of the desktop computer GPU RTX 1080TI.
• Participated in AISlanders’ Show 2018 and CES 2019.


Emotion reading system 06 2016
I.I.S. Research, Framework:python/caffee
• Combined face detection and emotion recognition to build a speaker assistance system that captures
audience emotions in real time and provides feedback.

Video Stabilization[IIHMSP] 08 2014
Master’s Thesis, Framework:MATLAB
• Implemented SIFT feature matching to get the camera movement path and update it to a stable path with
content-preserving warping.
• Submitted to IIHMSP2014 and won the Excellent paper award. Full-Frame Video Stabilization via SIFT Feature Matching (Excellent Paper Awards)


High-Dynamic Range image mapping 05 2013
Senior project, Framework:MATLAB
• Developed a MATLAB-based HDR system using histogram equalization and entropy to map an HDR.

Camera Automatic Exposure and Automatic White Balance 09 2012
Senior project, Team Leader, Framework: quatus verilog on embedding system DE2-70
• Implemented verilog for an AE and AWB camera system on an FPGA-based embedded system.
• Led four students to participate in the FPGA contest held by Altera asia.

一月 2015 - 一月 2023