Computer Vision Lab Will Present Three Papers at ECCV 2024
- Date2024.09.04
- 4330
The Computer Vision Lab (CVL) from the Department of Computer Science (Supervisor: Professor Min Dong-bo) will be presenting three papers at “European Conference on Computer Vision (ECCV) 2024”.
In the rapidly evolving field of AI, presenting at international conferences often takes precedence over publishing in SCI journals in leading the latest research trends. Along with CVPR (Computer Vision and Pattern Recognition) and ICCV (International Conference on Computer Science), ECCV is recognized as one of the top international conferences in the computer vision and AI field, where key researchers, scholars, and industry experts gather to discuss the latest research and technology trends.
ECCV 2024 will be held in Milano from September 29th to October 4th. The CVL research team had three papers accepted for presentation, recognizing the academic importance and innovation of their research.
(From left) Professor Min, Choi Hye-song (Ph.D. student), Park Hye-jin (Ph.D. candidate)
The paper “Salience-Based Adaptive Masking: Revisiting Token Dynamics for Enhanced Pre-training” (Choi Hye-song (first author), Park Hye-jin, Lee Gwang-mu, Cha Seong-min, Min Dong-bo (corresponding author) is a collaboration with the Computer Vision Lab at the University of British Columbia (UBC) in Canada. It proposed a salience-based token masking technique and adaptive masking ratio for masked image modeling which is garnering attention in the field nowadays. This research significantly improves the performance and efficiency of pre-training, offering new directions for learning computer vision models.
Another paper, “Emerging Property of Masked Token for Effective Pre-training” (Choi Hye-song (first author), Lee Heon-sang, Jeong Sae-young, Park Hye-jin, Kim Ji-young, Min Dong-bo (corresponding author), collaborated with Hyundai Motor Company’s R&D team, analyzed the new properties of masked tokens, which play a vital role in large model pre-training. Moreover, by introducing an entropy-based optimization technique, the research successfully reduced pre-training time by half and maximized training efficiency, thereby gaining people’s attention.
The third paper, “Dynamic Guidance Adversarial Distillation with Enhanced Teacher Knowledge” (Park Hye-jin (first author), Min Dong-bo (corresponding author), proposed an innovative approach to overcoming the vulnerability of AI models against adversarial attacks. The DGAD (Dynamic Guidance Adversarial Distillation) framework was proposed to address the transfer of prediction errors, enabling student models to achieve high accuracy and robust defense capabilities simultaneously.
These achievements are expected to elevate the global status of the CVL and serve as a significant milestone in advancing research in the computer vision and AI field.