Published on: 18-Jul-2019
NTIRE (http://www.vision.ee.ethz.ch/ntire19/) is the largest and most important competition on image restoration and image enhancement. The competition this year was held in conjunction with CVPR 2019.
Associate Professor Loy Chen-Change (far left) led a joint team with members from The Chinese University of Hong Kong, Nanyang Technological University, Shenzhen Institutes of Advanced Technology, and SenseTime to develop a new deep learning method called EDVR. With the method, the team won all four tracks under the video restoration challenge, which included the tasks of video super-resolution and video deblurring.
It is noteworthy that this was the first video restoration challenge after three years of consecutive NTIRE competitions that focused on image restoration. In comparison to image restoration, video restoration is more challenging due to large and complex motions between frames. To cope with large motions, the team devises a Pyramid, Cascading and Deformable (PCD) alignment module, in which frame alignment is done at the feature level using deformable convolutions in a coarse-to-fine manner. In addition, the team proposes a Temporal and Spatial Attention (TSA) fusion module, in which attention is applied both temporally and spatially, so as to emphasize important features for subsequent restoration.
Thanks to these modules, the proposed EDVR placed first and significantly ahead of its competitors in terms of speed and restoration quality.
The team has released the paper entitled “EDVR: Video Restoration with Enhanced Deformable Convolutional Networks” on arXiv.
The source code is available at GitHub:
Team members: Xintao Wang (CUHK), Kelvin C. K. Chan (NTU), Ke Yu (CUHK), Chao Dong (SIAT), Chen-Change Loy (NTU)
The team would like to thank SenseTime for their generous support on GPU resources.
Back to listing