M.S., Graduate Institute of Communication Engineering
National Taiwan University | Sep. 2023 - Jun. 2025
Advisor: Soo-Chang Pei, and Jian-Jiun Ding
GPA: 4.2 / 4.3
Master's Thesis: "Image Segmentation and Depth Estimation in Multi-Light Underwater Scenes: Environmental Adaptation and Robust Vision Methods"
B.S., Department of Electrical Engineering
National Chung Cheng University | Sep. 2019 - Jun. 2023
Honor Award: Awarded for five semesters
Dean's List Award: One semester
Publication
Boundary-Aware Refinement with Environment-Robust Adapter Tuning for Underwater Instance Segmentation
| Pin-Chi Pan, and Soo-Chang Pei
Submitted to the 17th Asian Conference on Machine Learning (ACML 2025 under review)
Underwater instance segmentation is challenged by light attenuation, scattering, and color distortion.
We propose BARD-ERA, a unified framework with BARDecoder for progressive boundary refinement,
ERA for efficient adaptation to degradations with over 90% fewer parameters,
and BACE loss for stronger boundary supervision.
UWSegDepth: Semantic-Aware Object-Level Depth Estimation in Underwater Scenes
| Pin-Chi Pan, and Soo-Chang Pei
The 38th Conference on Computer Vision, Graphics, and Image Processing (CVGIP 2025)
We propose Segmentation-Augmented Differential Depth Estimation Regressor (SADDER),
a lightweight module leveraging instance segmentation to correct residual errors,
and UWSegDepth, a post-processing method that averages depths per segmented object to enhance object-level spatial structure.
Global-Local Awareness Network for Image Super-Resolution
| Pin-Chi Pan, Tzu-Hao Hsu, Wen-Li Wei, and Jen-Chun Lin
2023 IEEE International Conference on Image Processing (ICIP 2023)
While self-attention excels at modeling global information,
it is less effective at capturing high frequencies (e.g., edges etc.)
that deliver local information primarily, which is crucial for SISR. To tackle this,
we propose a global-local awareness network (GLA-Net) to effectively capture global and local information
to learn comprehensive features with low- and high-frequency information.
LogoGANs: Generating and Compositing Multimodal Logo based on Generative Adversarial Networks
| Pin-Chi Pan, and Alan Liu
Taiwanese Association for Artificial Intelligence, 2022 (TAAI 2022)
This research endeavors to introduce a novel architecture of generative adversarial network aimed at producing multimodal logos.
Our focus lies in enhancing the model performance of compositional generative adversarial networks by integrating them with spatial transformation networks.
Work Experience
Sep. 2025 - Present
Taipei, Taiwan
Algorithm Engineer
Ganzin Technology
Analyze, design, and implement computer vision algorithms for biometric eye tracking with emphasis on performance, efficiency, and flexibility.
Collaborate closely with hardware and software teams to integrate algorithms into AR/VR/MR devices.
Sep. 2022 - Sep. 2024
Taipei, Taiwan
Research Assistant
Institute of Information Science, Academia Sinica
Conducted research on low-level vision inverse problems (e.g., image restoration).
Contributed to model development and co-authored international publications.
Jul. 2022 - Aug. 2022
Taipei, Taiwan
Summer Intern
Institute of Information Science, Academia Sinica
Conducted research on image super-resolution within low-level vision tasks.
Contributed to model refinement and evaluation of experimental outcomes.
Mar. 2022 - Sep. 2022
Chiayi, Taiwan
Website Developer
College Admissions Committee
Designed front-end pages and implemented back-end functionality.
Developed and tested websites based on user requirements.
Sep. 2021 - Sep. 2022
Chiayi, Taiwan
Website Developer
Office of Information Technology
Designed front-end pages and implemented back-end functionality.
Developed and tested websites based on user requirements.
Projects
Human Mesh Recovery with Optimization Guidance (HMROpt)
2024 CVPDL Final Project | Sep. 2024 - Dec. 2024
Proposed a novel framework integrating optimization with diffusion-based score guidance, achieving a 3.4 mm reduction in keypoint fitting error on 3DPW compared to ScoreHMR (CVPR 2024).
Set new benchmarks in multi-view refinement and motion recovery, with 28.6 mm PA-MPJPE on Human3.6M and 48.4 mm error on 3DPW.
DiffMusic: A Zero-shot Diffusion-Based Framework for Music Inverse Problem
2024 DeepMIR Final Project | Nov. 2024 - Dec. 2024
We proposed DiffMusic, a zero-shot diffusion-based framework designed to solve various music inverse problems.
Leverages pretrained models for zero-shot conditional generation, provide 5 operation to enable flexible music processing without extensive fine-tuning.
Memory Visual Query Localization from Correspondence with Fine-Grained Alignment (MemVQLoC-FGA)
2023 DLCV Final Project | Nov. 2023 - Dec. 2023
Proposed MemVQLoC-FGA by enhancing VQLoC (NeurIPS 2023) with recurrent memory mechanism and
fine-grained alignment, achieving a 4.6 stAP25 gain on the validation set 25 from 24.2 to 28.8.
Ranked 1st out of 13 teams in the DLCV Final Project Challenge.
Screen-Based Gaze Tracking Model
Independent Study | Sep. 2021 - Nov. 2022
Developed a Screen-Based Gaze Tracking Model that addresses errors caused by head posturevariations in non-wearable devices during gaze prediction.
Enhanced screen gaze area prediction accuracy by incorporating facial coordinates and the distancebetween the face and the camera into decision tree training parameters.
