Chin-Yang Lin | 林晉暘

I am a Master's student at the Institute of Data Science and Engineering, College of Computer Science, NYCU, co-advised by Prof. Yu-Lun Liu at CompPhoto Lab and Prof. Wei-Chen Chiu at Enriched Vision Applications Lab, focusing on research in computer vision, neural radiance fields (NeRF), and computational photography.

Previously, I completed my undergraduate studies at National Cheng Kung University, where I double-majored in Computer Science and Geomatics.

Email  /  CV  /  Google Scholar  /  LinkedIn  /  Github  /  YouTube

News

Mar, 2025	Two papers accepted by CVPR 2025
Dec, 2024	Serve as a student volunteer for SIGGRAPH Asia 2024
Oct, 2024	Serve as a reviewer for ICLR 2025
Jul, 2024	Start my imaging engineer internship at Logitech
Mar, 2024	One paper accepted by SIGGRAPH 2024

Research

I'm interested in computer vision, deep learning, generative AI, and image processing, focusing on efficient algorithms for novel view synthesis, camera pose estimation, and sensor fusion.

FrugalNeRF turns just two images and 10 minutes into high-quality 3D scenes by using weight-sharing voxels and cross-scale geometric adaptation to guide training without relying on external priors.

A reference-guided 3D inpainting approach utilizing SDEdit on aligned Gaussian initialization, and created a 360° inpainting dataset (360-USID) for comprehensive evaluation.

DiffIR2VR-Zero leverages pre-trained diffusion models for video restoration, using hierarchical token merging and hybrid optical flow with nearest neighbor matching. It achieves top performance across diverse degradations without training.

Our method does not require training and can adapt to any MVS-based NeRF methods in a feed-forward fashion to improve rendering quality.

We proposed a GNSS/PDR fusion algorithm specifically designed for smartwatches. This algorithm tracks the varied roll and pitch of the sensor caused by hand swings and integrates a CNN model to predict 1-D speed and perform ZUPT detection.

Projects

CUDA-ISP (Image Signal Processing) is a high-performance image processing pipeline designed for CUDA-enabled GPUs.

Boosting zero-shot CLIP segmentation with a fast bilateral solver enables accurate estimation of water regions while preserving fine boundaries. Achieving up to 0.90 IoU on the test set surpasses other methods involving training or fine-tuning.

Achieving fast convergence of high-quality NeRF with only two input images by utilizing voxel representation and integrating visibility priors and monocular depth, reducing the training time by 30x.

Stereo camera navigation in point cloud maps using ICP registration in challenging scenarios such as underground parking lots.

A wrist-worn IMU PDR algorithm. Utilizing VQF to compute IMU attitude, divided into three stages: Step Detection, Step Length Estimation, and Heading Estimation, enabling navigation with Wrist-Worn IMU worn by pedestrians.

A ROS package for real-time monocular camera visual odometry based on OpenCV functions, including feature matching and camera pose estimation.

A ROS package for robot navigation and arm control. It is capable of recognizing and grasping objects, as well as navigating to any location on the map.

A real-time camera navigation algorithm in a pre-built LiDAR map, utilizing NDT for 3D point cloud registration, effectively reducing 65% of accumulated position error.

Design and source code from Jon Barron's website.

Last updated Mar 2025.