Kun-Yu Lin

I am now a post-doctoral research fellow at the University of Hong Kong, under the supervision of Prof. Kai Han. I earned my PhD degree from Sun Yat-sen University, under the supervision of Prof. Wei-Shi Zheng. Before that, I received my Bachelor's degree and Master's degree from Sun Yat-sen University. During my PhD, I was fortunate to have the opportunity to study as a visiting student at MMLab@NTU, under the supervision of Prof. Chen Change Loy and Prof. Henghui Ding. My research interests include computer vision and machine learning.

Email  /  Scholar  /  GitHub

News

Selected Works

Most of my research works are about video understanding, trustworthy deep learning, and vision-language models. Some representative works are highlighted. # denotes equal contributions. ✉ denotes corresponding author.

A novel vision-language task, panoptic captioning, that produces comprehensive text representations toward the conceptual minimum text equivalent of images.

A novel 3D dense captioning task, holo-captioning, that produces comprehensive structured text representations toward the text equivalent of 3D scenes. Panoptic captioning and holo-captioning are complementary steps toward our ultimate goal of visual-text equivalence.

A strong referring video object segmentation model based on visual grounding foundations, serving as the core of the runner-up solution for the PVUW Challenge RVOS Track at CVPR 2025.

A cross-task manipulation generalization benchmark to evaluate existing Vision-Language-Action (VLA) models, together with a generalizable VLA method.

A new paradigm that uses paired human-robot videos to adapt human-data-pretrained models for robotic manipulation.

A general unlearning solution for any class-centric tasks, requiring neither retained data nor pretrained model knowledge.

A novel connector that bridges vision and language modalities by leveraging both global and partial views, along with a large-scale image-text dataset of detailed captions.

A novel task for studying the generalizability of open-vocabulary action recognition models, along with a strong method for improving generalization.

A human-centric video network to address the context bias in domain adaptive action recognition.

A diversity-aware video network to address domain-specific bias in video domain generalization.

A new event-guided paradigm to address the semantic gap between observed states and unobserved actions for procedure planning in instructional videos.

An asymmetric adaptation paradigm for few-shot domain adaptive object detection.

A new vision transformer architecture for efficient and effective visual understanding.

A theoretical method for balancing in-distribution generalization and out-of-distribution detection.

A theoretical method based on implicit generators to improve out-of-distribution detection.

A simple yet effective method based on cycle transformation to filter out outlier classes in partial domain adaptation.

Services

This website borrows from Jon Barron.