Abstract
Restoration of real-world analog video is a challenging task due to the presence of very heterogeneous defects. These defects are hard to model, such that creating training data synthetically is infeasible and instead time-consuming manual editing is required. In this work we explore whether reasonable restoration models can be learned from data without explicitly modeling the defects or manual editing. We adopt Noise2Noise techniques, which eliminate the need for ground truth targets by replacing them with corrupted instances. To compensate for temporal mismatches between the frames and ensure meaningful training, we apply motion correction. Our experiments show that video restoration can be learned using only corrupted frames, with performance exceeding that of conventional learning.
Original language | English |
---|---|
Title of host publication | Joint Austrian Computer Vision and Robotics Workshop 2020 |
Publisher | Verlag der Technischen Universität Graz |
Pages | 145-150 |
Number of pages | 5 |
ISBN (Electronic) | 978-3-85125-752-6 |
DOIs | |
Publication status | Published - 2020 |
Event | 2020 Joined Austrian Computer Vision and Robotics Workshop - Technische Universität Graz, abgesagt, Austria Duration: 17 Sept 2020 → 18 Sept 2020 |
Conference
Conference | 2020 Joined Austrian Computer Vision and Robotics Workshop |
---|---|
Abbreviated title | ACVRW '20 |
Country/Territory | Austria |
City | abgesagt |
Period | 17/09/20 → 18/09/20 |