Kenji Enomoto, Research Scientist, Adobe Research
Abstract: This talk delves into the progression of image and video matting techniques, focusing on the limitations of current object selection tools and the complexities inherent in the matting problem. I will explore recent advances in learning-based image matting and its extension to video, with a particular focus on the critical lack of high-quality video matting datasets. I will introduce PolarMatte, our innovative approach that utilizes polarization cues to generate ground-truth-quality alpha mattes for both images and videos, eliminating the need for manual labeling. This presentation will demonstrate how PolarMatte overcomes the scalability challenges in matting data collection, potentially transforming the field by enabling the creation of large-scale, high-quality matting datasets.
PolarMatte paper: https://openaccess.thecvf.com/content/CVPR2024/papers/Enomoto_PolarMatte_Fully_Computational_Ground-Truth-Quality_Alpha_Matte_Extraction_for_Images_and_CVPR_2024_paper.pdf
Bio: Kenji Enomoto is a research scientist at Adobe Research in San Jose, where he works on computer vision and machine learning. He is particularly interested in emerging devices that push the boundary of research. He received a PhD in Information Science from Osaka University, Japan, in 2022, where he was advised by Yasuyuki Matsushita.
https://enomotokenji.github.io/
https://www.meetup.com/sv-siggraph/events/305322896 For Zoom Info
0:00 Chapter Intro
2:53 Speaker Intro
3:55 Presentation
54:28 Q&A