Byeonggi Kwon and Hyunwoo Nam, doctoral students in POSTECH's Department of Electrical Engineering advised by Professor Tae-Hyun Oh, won Bronze in Signal Processing at Samsung Electronics' 29th Humantech Paper Awards. Including another Bronze in Computer Engineering for joint research with Professor Seokhyeong Kang's team, AMI Lab earned two Bronze Awards in this competition.

The winning paper, “DFlow: Learning to Synthesize Better Optical Flow Datasets via a Differentiable Pipeline,” proposes a new approach to optical flow estimation, a fundamental computer vision problem of estimating motion from camera images. Recent approaches use AI models requiring large training datasets. However, collecting data for this problem is difficult, and the largest publicly available real-world dataset contained only about 1,000 samples.

The team developed a system that learns from limited data what makes a dataset favorable for model training and generates synthetic data reflecting those criteria. It readily produces large datasets, substantially improves overall model performance, and enables robust optical-flow training even for snowy or rainy conditions. These achievements earned the Bronze Award. The research is scheduled for presentation in May at the leading International Conference on Learning Representations (ICLR) 2023 in Rwanda.

A separate department news article describes the Computer Engineering Bronze Award for “FPGA-Based Accelerator for Rank-Enhanced and Highly-Pruned Block-Circulant Neural Networks,” joint research by Oh's and Kang's teams led by Haena Song, advised by Kang.

Established by Samsung Electronics in 1994, the Humantech Paper Awards identify and develop outstanding science/technology talent to secure world-class competitiveness and technological capabilities. Samsung hosts the competition with joint support from the Ministry of Science and ICT and JoongAng Ilbo. The authors were selected through a rigorous three-stage review from 1,972 papers at the 29th competition.