AI Researcher at UNIST AIGS

Junsu Kim

Continual Learning Open-world Detection Vision-Language + Diffusion
Current Base UNIST Busan Centum Campus

Can AI learn without forgetting?

I study how vision models keep learning as the world changes, using continual learning, vision-language models, and diffusion.

SDDGR · CVPR HighlightAll papers
Background & collaborators

I am a Combined MS/PhD student at UNIST AIGS, advised by Prof. Seungryul Baek. My work focuses on continual learning, open-world detection, and vision-language systems.

Previously, I interned at NAVER AI Lab advised by Dongyoon Han, where I studied backbone architectures and analyzed internal representations of large-scale models. I also serve as an NVIDIA AI Ambassador, with a strong interest in applied AI for everyday problems.

Email Scholar GitHub LinkedIn Blog CV
Junsu Kim smiling by the river with a city skyline at night
t=480
click to denoise
Recent Highlights
2026 Presented BeeHunters at CVPPA @ ECCV 2026.
2026 1st Place in the CVPPA@ECCV'26 BuzzSpot Challenge (Solo Entry · Small Object Detection) Code
2025 Outstanding Reviewer at BMVC 2025
2025 Jul 2025 · One paper accepted to BMVC (co-first author)
2025 Jul 2025 · Two papers accepted to ICCV Workshops (1st author)
Highlight CVPR 2024 first-author paper selected as Highlight Poster (top 2.8%)

Publications

0+ Total Citations

Showing 8 papers (+6 domestic)

BeeHunters tiny pollinator detection figure
ECCV 2026 Challenge · Technical Report
Where Is the Bee? Detecting Tiny Pollinators with a Single Collaborative-Head Transformer
J. Kim, S. Baek
Paper Code
BeeHunters project figure
A single-model, first-place BuzzSpot Challenge solution for tiny pollinator detection, combining Co-DINO with a Swin-L backbone, crop-mosaic training, and class-weighted ETF supervision.
B-RIGHT figure
BMVC 2025
B-RIGHT: Benchmark Re-evaluation for Integrity in Generalized Human-Object Interaction Testing
J. Kim*, Y. Jang*, H. Kim, E. Lee, E. Kim, S. Baek, J. Yoo
Paper
B-RIGHT Figure
A systematic re-evaluation of existing HOI benchmarks, revealing integrity issues and proposing new evaluation protocols for generalized human-object interaction testing.
FSCIL Synthetic figure
ICCV-W 2025 4 citations
Can Synthetic Images Conquer Forgetting? Beyond Unexplored Doubts in Few-Shot Class-Incremental Learning
J. Kim, Y. Ku, S. Baek
Paper
FSCIL Synthetic Figure
Investigates whether synthetic image generation can effectively address catastrophic forgetting in few-shot class-incremental learning, challenging common assumptions in the field.
RPE figure
ICCV-W 2025
Revisiting Reliability in the Reasoning-based Pose Estimation Benchmark
J. Kim, N. Kim, J. Lee, I. Park, D. Han, S. Baek
Paper Code & Data
RPE Figure
Audits the RPE benchmark, uncovers annotation and image-index mismatches that undermine reproducibility, and releases corrected ground-truth annotations with evaluation code.
Diffusion FSCIL figure
ArXiv 2025 2 citations
Beyond Synthetic Replays: Turning Diffusion Features into Few-Shot Class-Incremental Learning Knowledge
J. Kim, Y. Ku, D. Han, S. Baek
Paper
Diffusion FSCIL Figure
Bridges diffusion models with few-shot class incremental learning, leveraging generative priors to overcome data scarcity and catastrophic forgetting simultaneously.
SDDGR figure
CVPR Highlight · Top 2.8%
CVPR 2024 · Highlight 98 citations
SDDGR: Stable Diffusion-based Deep Generative Replay for Class Incremental Object Detection
J. Kim, H. Cho, J. Kim, Y. Tiruneh, S. Baek
Paper
SDDGR Figure
Proposes using Stable Diffusion as a generative replay mechanism for class incremental object detection, effectively mitigating catastrophic forgetting while maintaining high-quality synthetic data generation.
VLM-PL figure
CVPR-W 2024 54 citations
VLM-PL: Advanced Pseudo Labeling Approach for Class Incremental Object Detection with Vision-Language Model
J. Kim, Y. Ku, J. Kim, J. Cha, S. Baek
Paper
VLM-PL Figure
Leverages vision-language models for advanced pseudo labeling in class incremental object detection, improving detection accuracy on novel classes without forgetting.
ICASSP figure
ICASSP 2024 5 citations
Class-Wise Buffer Management for Incremental Object Detection: An Effective Buffer Training Strategy
J. Kim, S. Hong, C. Kim, J. Kim, Y. Tiruneh, J. On, J. Song, S. Choi, S. Baek
Paper
ICASSP Figure
Introduces a class-wise buffer management strategy for incremental object detection that optimizes exemplar selection and training to balance old and new class performance.
IEIE Summer 2023
Reducing Data Imbalance for Object Detection
J. Kim, S. Hong, C. Kim, J. Kim, J. On, Y. Tiruneh, J. Song, S. Choi, S. Baek
JKAIA 2022
Temporal MANO-Based Video Hand Pose and Shape Estimation
H. Park, J. Kim, J. Kim, S. Baek
Best Paper Award
IEIE Fall 2021
Multi-Task Learning Approach to Autonomous Driving Techniques
I. Lee, J. Kim, H. Lee, Y. Kim, D. Lee, H. Lee
Best Paper Award
KICS Summer 2021
Deep Learning-Based Self-Driving Systems for Swarm Vehicles
H. Kim, J. Kim, I. Lee, H. Lee
Best Paper Award
IEMEK ICT 2021
AI Edge Server Design for Swarm Vehicular Networks
J. Kim, H. Kim, I. Lee, H. Lee
KICS Winter 2021
AI-Enabled Low-Complexity Autonomous Driving Using TensorRT
H. Kim, J. Kim, I. Lee, H. Lee

Projects

5 projects
BeeHunters tiny pollinator detection project preview
Challenge winner

BeeHunters

The 1st-place single-model solution to the CVPPA@ECCV'26 BuzzSpot Challenge, detecting tiny pollinators with Co-DINO, Swin-L, and a collaborative-head ETF objective.

Small objects Co-DINO 1st Place
Paper GitHub Results
SparkOrbit project preview
AI dashboard

SparkOrbit

AI orbit monitor for papers, models, benchmarks, and company news in one dashboard, with 40+ sources and local LLM summaries.

AI monitor Dashboard Local LLM
GitHub
MOMO project preview
Meeting intelligence

MOMO

A local-first pipeline that turns long Zoom recordings into Markdown meeting recaps, keeping a separate evidence file for every claim.

Python Whisper Ollama
GitHub
Cheapest Flights project preview
Travel search

Cheapest Flights

A TypeScript tool that sweeps departure and return date combinations, then ranks cheap direct-flight candidates across travel sites.

TypeScript Flight search Date sweep
GitHub
LOOP-STATION project preview
Agent workflow

LOOP-STATION

A Codex skill for bounded goal loops, multi-agent review, adaptive sessions, and implementation variants that stay controlled.

Codex skill Multi-agent Review loop
GitHub

Education

2023 – Present
Combined MS-PhD in AI
UNIST, Ulsan, South Korea
Advisor: Prof. Seungryul Baek
2016 – 2022
B.S. in Information and Communications Engineering
Pukyong National University, Busan
Research advised by Prof. Hoon Lee since 2019

Experience

2024 – 2025
AI Research Intern
NAVER AI Lab, Seoul
Backbone Research Team · Advisor: Dongyoon Han
2022 – 2025
NVIDIA AI Ambassador
NVIDIA DLI Program
Leader, NVIDIA · MODULABS Foundation Models Lab
2022 – 2023
AI Research Intern
UVL Lab, Ulsan & Corp. KST, Busan

Awards

2026 1st Place, CVPPA@ECCV'26 BuzzSpot Challenge (Solo Entry · Small Object Detection) Code
2025 Outstanding Reviewer, BMVC 2025
2025 Excellence Award, 37th IPIU Poster Presentation
2021 Best Paper Award, IEMEK · KICS · IEIE
2021 1st Prize, PKNU Capstone Design
2020 2nd Prize, 기술보증기금 AI Contest

Honored To

2025 Serve as reviewer for AAAI, BMVC (Outstanding), IEEE TMM, IEEE TIP
2025 Be invited as Young Researcher to National Physical-AI Strategy Advisory Seminar (Invited by ETRI)
2024 Present oral talk at KOREA Digital Innovators Symposium as UNIST representative
2024 Present poster at KCCV 2024

Teaching & Mentoring

Samsung SSAFY (13th)

Lecture Mentor

Language Models & RAG

UNIST AI Novatus Academia

Workshop Lecturer

Computer Vision

NVIDIA Jetson AI Workshop

NVIDIA Instructor

Pukyong National University

BeeHunters · CVPPA @ ECCV 2026

Junsu Kim presenting the BeeHunters detector comparison at CVPPA, ECCV 2026
NVIDIA AI Ambassador
NVIDIA DLI Certified Instructor Directory profile screenshot