Featured
OTT-Vid: Optimal Transport Temporal Token Compression for Video Large Language Models
Under Review at NeurIPS 2026
An optimal transport based temporal token compression method for efficient video large language models.
Scene Understanding · Video Temporal Grounding
Vision-Language Multimodal Learning
Minseok Kang
Integrated M.S./Ph.D. Student in Electrical and Electronic Engineering at Yonsei University, working on computer vision and machine/deep learning for video.
My research centers on scene understanding and vision-language multimodal learning, including video temporal grounding, video scene graph generation, and video large language models. I am interested in building models that reason about motion, long-range temporal structure, and language for real-world video understanding.
01
About
Researcher Profile
I am an Integrated M.S./Ph.D. student in Electrical and Electronic Engineering at Yonsei University, in the Image and Video Pattern Recognition Lab (MVP Lab), advised by Prof. Sangyoun Lee. I received my B.S. in Electrical and Electronic Engineering from Korea University.
My research lies in computer vision and machine/deep learning, with a focus on scene understanding, video temporal grounding, and video scene graph generation, as well as vision-language multimodal learning for video large language models.
I am interested in models that can understand dynamic visual content, reason about motion and long-range temporal structure, and connect vision with language for robust real-world video understanding.
Affiliation
Yonsei University, MVP Lab
Advisor
Prof. Sangyoun Lee
Program
Integrated M.S./Ph.D. (2022 – 2027 expected)
Interests
Scene Understanding, Video Temporal Grounding, Vision-Language Multimodal Learning
02
News
Recent Updates
Jun 2026
ECCV 2026 Accepted
First-author paper "PAWS" accepted to ECCV 2026.
May 2026
NeurIPS 2026 Submission
First-author paper "OTT-Vid" under review at NeurIPS 2026.
Feb 2026
CVPR Findings 2026 Accepted
Two co-authored papers accepted to CVPR Findings 2026.
Feb 2026
Pattern Recognition 2026 Accepted
Co-authored paper on zero-shot anomaly detection accepted to Pattern Recognition.
Oct 2025
3rd Place — Perception Test Challenge 2025
3rd place in Task 3 (Temporal Action & Sound Localization), hosted by DeepMind at the ICCV 2025 Workshop.
Sep 2025
NeurIPS 2025 Accepted
First-author paper "DualGround" accepted to NeurIPS 2025.
May 2025
ICIP 2025 Accepted
Co-authored paper "CMTM" accepted to ICIP 2025.
03
Publications
Selected Publications
Featured
Revisiting Weakly-Supervised Video Scene Graph Generation via Pair Affinity Learning
ECCV 2026 (Accepted)
A pair affinity learning approach for weakly-supervised video scene graph generation.
Featured
Empower Words: DualGround for Structured Phrase and Sentence-Level Temporal Grounding
NeurIPS 2025 (Accepted)
A structured framework for phrase- and sentence-level temporal grounding in video.
SwiftVGGT: A Scalable Visual Geometry Grounded Transformer for Large-Scale Scenes
CVPR Findings 2026 (Accepted)
Seen-to-Scene: Keep the Seen, Generate the Unseen for Video Outpainting
CVPR Findings 2026 (Accepted)
Generalizing CLIP Prompts for Zero-Shot Anomaly Detection
Pattern Recognition 2026 (Accepted)
CMTM: Cross-Modal Token Modulation for Unsupervised Video Object Segmentation
ICIP 2025 (Accepted)
Multilevel Artificial Electronic Synaptic Device of Direct Grown Robust MoS2 Based Memristor Array for In-Memory Deep Neural Network
npj 2D Materials and Applications, 2022 (Accepted)
04
Research
Research Interests
Scene Understanding
Building models that parse complex visual scenes, reason about objects and their relations, and capture structured semantics in video.
Topics: Video Scene Graph Generation, Structured Reasoning
Video Temporal Grounding
Localizing moments and actions in video given natural language, with a focus on phrase- and sentence-level temporal grounding.
Topics: Temporal Grounding, Moment Retrieval, Action Localization
Vision-Language Multimodal Learning
Connecting vision and language to improve understanding and reasoning over dynamic visual content.
Topics: Video Large Language Models, Cross-Modal Learning
Efficient Video Understanding
Designing token compression and efficient architectures that make large-scale video models practical for real-world deployment.
Topics: Token Compression, Optimal Transport, Efficient Modeling
Robust Perception
Developing recognition systems that remain reliable under low-visibility, adverse weather, and multi-sensor conditions.
Topics: Multi-Sensor Fusion, Robust Detection, Domain Generalization
Video Object Segmentation
Studying motion-guided and cross-modal methods for segmenting salient objects in video, including unsupervised settings.
Topics: Unsupervised VOS, Motion Cues, Cross-Modal Modulation
Contact
Let's Connect
I am always happy to discuss research and collaboration.
please feel free to get in touch.