Modeling Visual Data Structures for Semantic Segmentation

Talk by Tianfei ZHOU

Sep 14, 2023 Thursday

Abstract:

Semantic segmentation aims at predicting one semantic category for each pixel in an image. It is a vital step towards intelligent scene understanding. With the renaissance of connectionism, rapid progress has been made in the field. However, the vast majority of modern segmentation models is pure data-driven, and ignores the structured nature of visual data. In this talk, I will first introduce a pixel-wise metric learning paradigm for semantic segmentation, which explicitly explores cross-image pixel relations in large-scale training dataset for semantic segmentation learning. Moreover, I will present our investigations towards hierarchy-aware segmentation, which aims to better reflect the structured nature of our visual world and echo the hierarchical reasoning mode of human visual perception.

Time:

September 14th, 2023, Thursday

16:30 – 17:20

Location:

Rm101, E1

Zoom:

628 334 1826 (PW: 234567)

Bilibili Live:

ID: 30748067

Speaker Bio:

Dr. Tianfei Zhou

Researcher, Computer Vision Lab, ETH

Dr. Tianfei Zhou is currently a researcher at Computer Vision Lab, ETH Zurich. He obtained his Ph.D. degree from Beijing Institute of Technology and has worked as a researcher at Lenovo Research and Inception Institute of Artificial Intelligence. His general research interests are computer vision and machine learning, and specifically his recent focus is on building customizable multi-modal large models that can understand and follow humans’ intent. He has published over 50 academic papers in top-tier journals conferences, such as IEEE TPAMI, ICML, ICLR, CVPR, and MICCAI. As the first authorship, he received the Best Paper Award at MICCAI 2022. He has won 5 champions in international academic competitions in conjunction with, e.g., CVPR and ICCV. He was a Guest Editor at IEEE TCSVT.