1.Title: StyleGAN series and their applications in image generation and manipulation
Speaker: Anh Tran is an AI Research Scientist in VinAI Research. Before that, he was a Software Developer Engineer and an Applied Scientist at Amazon AWS Rekognition. He took a PhD degree from Computer Science Department of University of Southern California, under the support of the Vietnam Education Foundation funded by the U.S. government. His advisor was Prof. Gerard Medioni. Previously, he received a BSc in Computer Science from Hanoi University of Science and Technology. He received the first prizes in Vietnam Talent Award 2010 and Imagine Cup Vietnam 2009. His research interest is Computer Vision, and his expertise is Facial Image Processing. He has about 20 papers accepted to top conferences in Computer Vision and Machine Learning. He also served as an Area Chair at ICCV 2021, CVPR 2022, and ECCV 2022
Abstract: This tutorial will introduce StyleGAN, the state-of-the-art high-resolution image generation model series developed by NVIDIA. These GAN models can synthesize diverse, realistic images from random noise vector inputs with a resolution up to 1024x1024. We will go through the structure and training procedure of the first version (StyleGAN), the modifications in the second version (StyleGAN2), the augmentation-based solution for training with limited data (StyleGAN2-Ada), and the latest version with equivariant image synthesis (StyleGAN3). We then discuss an application of StyleGAN in image manipulation with two main steps: GAN inversion and latent space traversal. We survey the state-of-the-art techniques to resolve the problems in each step. Finally, we will discuss the application of StyleGAN structure in other domains, such as generative novel view synthesis
2.Title: Few-shot object segmentation in images
Speaker: Dr. Khoi Nguyen is a Research Scientist at VinAI which is the Top-1 AI research company in Vietnam and ranked 20th in the world. His research interests include computer vision and machine learning with the focus on 3D reconstruction, object detection and segmentation, few-shot learning, and unsupervised learning. He has published many papers on object segmentation and few-shot learning in various domains including images, videos, and 3D point clouds at top-tier AI conferences such as NeurIPS, CVPR, ICCV, and ECCV. He also facilitates the research community by serving as a reviewer at top-tier conferences and journals in computer visions and machine learning and received the outstanding reviewer award from ECCV 2020.
Abstract: This tutorial will be about few-shot object segmentation in images -- a fundamental computer vision problem arising in many applications with access to only a few examples of target object classes due to, e.g., their rarity and the cost of labeling pixel-wise annotation. In this tutorial, we will go through three few-shot object segmentation tasks, namely few-shot semantic segmentation, few-shot instance segmentation, and incremental few-shot instance segmentation addressed by our three accepted papers in top-tier computer vision conferences. Our main research accomplishments in this area will be presented, including: explicit modeling and discovery of latent object parts shared across object classes; data uncertainty map estimation for each instance segmentation and data uncertainty based regularization of our few-shot learning; and model uncertainty estimation with an efficient approximation based on the probit function. Our contributions produce statistically significant performance gains over the state of the art on the benchmark COCO dataset