Siniša Šegvić completed his PhD degree in 2004 at the University of Zagreb. He spent one year as a postdoc researcher at IRISA Rennes (2006) and another year as a postdoc researcher at TU Graz (2007). Currently he is an associate professor at the University of Zagreb. He is a program committee member of the VISAPP 2018 conference and an associate editor of the Journal of Computing and Information Technology. His research expertise is in the fields of computer vision and deep learning, with special interest in applications for autonomous vehicles and safe traffic. He published more than 50 national and international scientific papers.
Lightweight convolutional models for real-time dense prediction and forecasting
Convolutional models have become an indispensable ingredient for automated visual recognition. Most of the existing work in the field is based on heavyweight models with hundreds of layers. Unfortunately, such designs are unable to offer real-time performance on embedded hardware, which precludes many important applications such as autonomous driving. However, recent work shows that convolutional models are extremely resistant to overfitting, contrary to established theories of machine learning.
This suggests that useful performance could be obtained with much less computational effort. Our latest findings support this notion in the fields of semantic segmentation prediction and forecasting. In particular, we demonstrate that the model capacity can be effectively compensated by careful wiring of information data paths within the model. Our approaches achieve state-of-the-art ratio between recognition accuracy and processing time, and show that real-time applications of convolutional models are indeed feasible.