List of participants of MSU Video Super Resolution Benchmark 2021

D3Dnet

Use a deformable 3D convolution to compensate the motion between frames implicitly.

Added to the benchmark by MSU

Links: GitHub, paper

DBVSR

Estimate a motion blur for the particular input. Compensate the motion between frames explicitly.

Added to the benchmark by MSU

Links: GitHub, paper

DUF

Two models: DUF-16L (16 layers), DUF-28L (28 layers)

Use a deformable 3D convolution to compensate the motion between frames implicitly.

Added to the benchmark by MSU

Links: GitHub, paper

DynaVSR

Two models: DynaVSR-R (trained on REDS), DynaVSR-V (trained on Vimeo-90k)

Use meta-learning to estimate a degradation kernel for the particular input.

DynaVSR can be applied to any VSR deep-learning model. For our benchmark, we used pretrained weights for model EDVR, which use Deformable convolution to align neighboring frames.

Added to the benchmark by MSU

Links: GitHub, paper

ESPCN

Added to the benchmark by MSU

Links: GitHub, paper

ESRGAN

Added to the benchmark by MSU

Links: GitHub, paper

iSeeBetter

Added to the benchmark by MSU

Links: GitHub, paper

LGFN

Use deformable convolutions with decreased multi-dilation convolution units (DMDCUs) to align frames explicitly. Fuse features from local and global fusion modules.

Added to the benchmark by MSU

Links: GitHub, paper

RBPN

Added to the benchmark by MSU

Links: GitHub, paper

Real-ESRGAN

Added to the benchmark by MSU

Links: GitHub, paper

Real-ESRNet

Added to the benchmark by MSU

Links: GitHub, paper

RealSR

Try to estimate degradation kernel and noise distribution for better visual quality.

Added to the benchmark by MSU

Links: GitHub, GitHub, paper

RRN

Two models: RRN-5L (five residual blocks), RRN-10L (ten residual blocks)

Use recurrent strategy with sets of residual blocks to store information from previous frames.

Added to the benchmark by MSU

Links: GitHub, paper

RSDN

It divides the input into structure and detail components which are fed to a recurrent unit composed of several proposed two-stream structure-detail blocks.

Added to the benchmark by MSU

Links: GitHub, paper

SOF-VSR

Two models: SOF-VSR-BD (trained on gauss degradation type), SOF-VSR-BI (trained on bicubic degradation type)

Compensate motion by high-resolution optical flow, estimated from the low-resolution one in a coarse-to-fine manner.

Added to the benchmark by MSU

Links: GitHub, paper

TDAN

Use a deformable 3D convolution to compensate the motion between frames implicitly.

Added to the benchmark by MSU

Links: GitHub, paper

TGA

The input sequence is reorganized into several groups of subsequences with different frame rates. The grouping allows to extract spatio-temporal information in a hierarchical manner, which is followed by an intra-group fusion module and inter-group fusion module.

Added to the benchmark by MSU

Links: GitHub, paper

TMNet

Temporal Modulation Network was trained for Space-Time Video Super Resolution. The temporal information is integrated by deformable convolution with the multi-frame input.

Added to the benchmark by the author, Gang Xu.

Links: GitHub, paper