Hyolim Kang
I am a Ph.D. student studying computer science at Yonsei University, where I am advised by Professor Seon Joo Kim.
My primary research areas are computer vision and deep learning, with specific interests in video understanding, perception for action, vision for robotics, and representation learning.
I am actively looking for internship positions!
Education
Ph.D candidate in Computer Science, Yonsei University, Seoul, Korea [Mar 2020 - Feb 2025 (expected)] Advisor: Seon Joo Kim
B.S in Computer Science, Yonsei University, Seoul, Korea [Mar 2014 - Feb 2020]
GPA 3.85/4.3 -- Top 8% in Dept. of Computer Science
Research Objective
My main research direction revolves around the challenge of efficient streaming perception in embodied AI and robotics. This area is vital because, unlike traditional models, embodied agents often lack complete observational data, necessitating real-time, sequential decision-making based on partial, streaming inputs. This mirrors how humans process visual information continuously, not in isolated instances.
I am currently excited about developing a "generic visual streaming encoder" adaptable for various streaming perception tasks in computer vision. I question the current practice of independent tokenization, where short clips are processed independently in pretrained clip encoders. My aim is to develop a contextual video clip embedding model that adaptively extracts clip features within their context, using architectures like RNN or S4. This approach mirrors the evolution in NLP from static word embeddings like word2vec to dynamic, context-aware models such as BERT and ELMo.
You can assess my full research statement here.
Publications
ActionSwitch: Class-agnostic Detection of Simultaneous Actions in Streaming Videos
Hyolim Kang, Jeongseok Hyun, Joungbin An, Youngjae Yu, Seon Joo Kim
ECCV 2024
Object Aware Egocentric Online Action Detection
Joungbin An, Yunsu Park, Hyolim Kang, Seon Joo Kim
CVPR 2024 First Joint Egocentric Vision Workshop
MiniROAD: Minimal RNN Framework for Online Action Detection
Joungbin An, Hyolim Kang, Su Ho Han, Ming-Hsuan Yang, Seon Joo Kim
ICCV 2023
Project Page / Paper / Code
Soft-Landing Strategy for Alleviating the Task Discrepancy Problem in Temporal Action Localization Tasks
Hyolim Kang, Hanjung Kim, Joungbin An, Minsu Cho, Seon Joo Kim
CVPR 2023
ComMU: Dataset for Combinatorial Music Generation
Lee Hyun*, Taehyun Kim*, Hyolim Kang, Minjoo Ki, Hyeonchan Hwang, Kwanho Park, Sharang Han, Seon Joo Kim
Neurips 2022 Datasets & Benchmarks
UBoCo : Unsupervised Boundary Contrastive Learning for Generic Event Boundary Detection
Hyolin Kang*, Jinwoo Kim*, Taehyun Kim, Seon Joo Kim
CVPR 2022
CAG-QIL: Context-Aware Actionness Grouping via Q Imitation Learning
for Online Temporal Action Localization
Hyolin Kang, Kyungmin Kim, Yumin Ko, Seon Joo Kim
ICCV 2021
Selected as the best vision paper in the 2022 Yonsei AI Workshop
($2000 worth of prize)
Winning the CVPR'2021 Kinetics-GEBD Challenge: Contrastive Learning Approach
Hyolin Kang, Jinwoo Kim, Kyungmin Kim, Taehyun Kim, Seon Joo Kim
CVPR 2021 LOng-form VidEo Understanding (LOVEU) Workshop
1st place in both track1.1 & track1.2
About me
I'm also an amateur composer and piano player. If you are interested, here are some of my original pieces & playing!
Original Compositions (including a little fugue)
Playing Bach & Chopin