Professor Cho Sung-in's Research Team Has a Regular Paper Accepted at the Top International Conference NeurIPS (The Thirty-Ninth Annual Conference on Neural Information Processing Systems) 2025

▶ (From left) Professor Cho Sung-in; Han Sang-hoon, Nam Hyuk and Lee Sun-ho, master's students

The paper 'Doodle to Detect: A Goofy but Powerful Approach to Skeleton-based Hand Gesture Recognition', written by master's students Han Sang-hoon (first author), Lee Sun-ho (joint first author) and Nam Hyuk (joint first author) of Professor Cho Sung-in's research team (Smart Vision & Media Laboratory, SVM) with Professor Cho Sung-in (corresponding author), has been accepted for publication at NeurIPS (The Thirty-Ninth Annual Conference on Neural Information Processing Systems) 2025.

▶ Overview of Skeleton Kinematics Extraction Through Coordinated grapH (SKETCH)

SKETCH, presented in this research, offers a new way of resolving the problems that existing hand gesture recognition methods face from viewpoint and individual variations. Using per-axis min–max normalization and a dynamic range embedding (DRE) technique, the team enabled stable and accurate recognition regardless of the position and orientation of the gesture. The Skeleton Kinematics Extraction Through Coordinated grapH (SKETCH) module converts 3D hand joint data into graph images, enabling powerful feature extraction, and uses a Swin Transformer as the image classification model to improve gesture recognition accuracy.

The team verified the performance of the SKETCH module in experiments on the SHREC'19 and SHREC'22 datasets. The results recorded a higher detection rate (DR) and a lower false positive rate (FP) than existing hand gesture recognition methods, while achieving real-time processing at more than 30 FPS. This demonstrates that the method overcomes the domain gap problem faced by existing gesture recognition systems and can deliver stable performance in real-time settings.

The research is significant in that it improves the accuracy of hand gesture recognition while securing real-time processing performance. The team plans to continue developing the technology so that the SKETCH module can be used in a variety of real-world applications. It is expected in particular to establish itself as an important technology for improving human–computer interaction in intelligent robots and virtual reality (VR)/augmented reality (AR) environments.

The 39th Annual Conference on Neural Information Processing Systems (NeurIPS '25) will be held in San Diego, United States (2–7 December) and in Mexico City (30 November – 5 December). NeurIPS is a large international conference at the intersection of machine learning, artificial intelligence, computational neuroscience and other fields, and is listed among the outstanding conferences in the software field of the Korean Institute of Information Scientists and Engineers and among the outstanding international conferences in computer science under the BK21 Plus programme (BK IF=4).

References:

NeurIPS 2025 homepage: https://neurips.cc/

Paper link: https://neurips.cc/virtual/2025/poster/115527

Homepage of Professor Cho Sung-in's research team: https://sites.google.com/view/csi2267svm/