Model Zoo and Baselines

September 8, 2020 · View on GitHub

English | 简体中文

We provide:

  1. Official models converted directly from official released models
  2. Reproduced models with BasicSR. Pre-trained models and log examples are provided

You can put the downloaded models in the experiments/pretrained_models folder.

Download official pre-trained models

You can use the scrip to download pre-trained models from Google Drive.

python scripts/download_pretrained_models.py --method ESRGAN
# method can be ESRGAN, EDVR, StyleGAN, EDSR, DUF, DFDNet, dlib

Download reproduced models and logs

In addition, we upload the training process and curves in wandb.

Training curves in wandb

Contents

  1. Image Super-Resolution
    1. Image SR Official Models
    2. Image SR Reproduced Models
  2. Video Super-Resolution

Image Super-Resolution

When evaluation:

  • We crop scale border pixels in each border
  • Evaluated on RGB channels

Image SR Official Models

Exp NameSet5 (PSNR/SSIM)Set14 (PSNR/SSIM)DIV2K100 (PSNR/SSIM)
EDSR_Mx2_f64b16_DIV2K_official-3ba7b08635.7768 / 0.944231.4966 / 0.893934.6291 / 0.9373
EDSR_Mx3_f64b16_DIV2K_official-6908f88a32.3597 / 0.90328.3932 / 0.809630.9438 / 0.8737
EDSR_Mx4_f64b16_DIV2K_official-0c28773330.1821 / 0.864126.7528 / 0.743228.9679 / 0.8183
EDSR_Lx2_f256b32_DIV2K_official-be38e77d35.9979 / 0.945431.8583 / 0.897135.0495 / 0.9407
EDSR_Lx3_f256b32_DIV2K_official-3660f70d32.643 / 0.90628.644 / 0.815231.28 / 0.8798
EDSR_Lx4_f256b32_DIV2K_official-76ee1c8f30.5499 / 0.870127.0011 / 0.750929.277 / 0.8266

Image SR Reproduced Models

Experiment name conventions are in Config.md.

Exp NameSet5 (PSNR/SSIM)Set14 (PSNR/SSIM)DIV2K100 (PSNR/SSIM)
001_MSRResNet_x4_f64b16_DIV2K_1000k_B16G1_wandb30.2468 / 0.865126.7817 / 0.745128.9967 / 0.8195
002_MSRResNet_x2_f64b16_DIV2K_1000k_B16G1_001pretrain_wandb35.7483 / 0.944231.5403 / 0.893734.6699 / 0.9377
003_MSRResNet_x3_f64b16_DIV2K_1000k_B16G1_001pretrain_wandb32.4038 / 0.903228.4418 / 0.810630.9726 / 0.8743
004_MSRGAN_x4_f64b16_DIV2K_400k_B16G1_wandb28.0158 / 0.808724.7474 / 0.662326.6504 / 0.7462
201_EDSR_Mx2_f64b16_DIV2K_300k_B16G1_wandb35.7395 / 0.94431.4348 / 0.893434.5798 / 0.937
202_EDSR_Mx3_f64b16_DIV2K_300k_B16G1_201pretrain_wandb32.315 / 0.902628.3866 / 0.808830.9095 / 0.8731
203_EDSR_Mx4_f64b16_DIV2K_300k_B16G1_201pretrain_wandb30.1726 / 0.864126.721 / 0.74328.9506 / 0.818
204_EDSR_Lx2_f256b32_DIV2K_300k_B16G1_wandb35.9792 / 0.945331.7284 / 0.895934.9544 / 0.9399
205_EDSR_Lx3_f256b32_DIV2K_300k_B16G1_204pretrain_wandb32.6467 / 0.905728.6859 / 0.815231.2664 / 0.8793
206_EDSR_Lx4_f256b32_DIV2K_300k_B16G1_204pretrain_wandb30.4718 / 0.869526.9616 / 0.750229.2621 / 0.8265

Video Super-Resolution

Evaluation

In the evaluation, we include all the input frames and do not crop any border pixels unless otherwise stated.
We do not use the self-ensemble (flip testing) strategy and any other post-processing methods.

EDVR

Name convention
EDVR_(training dataset)_(track name)_(model complexity)

  • track name. There are four tracks in the NTIRE 2019 Challenges on Video Restoration and Enhancement:
    • SR: super-resolution with a fixed downsampling kernel (MATLAB bicubic downsampling kernel is frequently used). Most of the previous video SR methods focus on this setting.
    • SRblur: the inputs are also degraded with motion blur.
    • deblur: standard deblurring (motion blur).
    • deblurcomp: motion blur + video compression artifacts.
  • model complexity
    • L (Large): # of channels = 128, # of back residual blocks = 40. This setting is used in our competition submission.
    • M (Moderate): # of channels = 64, # of back residual blocks = 10.

Download Models from Google Drive

Model name[Test Set] PSNR/SSIM
EDVR_Vimeo90K_SR_L[Vid4] (Y1) 27.35/0.8264 [↓Results]
(RGB) 25.83/0.8077
EDVR_REDS_SR_M[REDS] (RGB) 30.53/0.8699 [↓Results]
EDVR_REDS_SR_L[REDS] (RGB) 31.09/0.8800 [↓Results]
EDVR_REDS_SRblur_L[REDS] (RGB) 28.88/0.8361 [↓Results]
EDVR_REDS_deblur_L[REDS] (RGB) 34.80/0.9487 [↓Results]
EDVR_REDS_deblurcomp_L[REDS] (RGB) 30.24/0.8567 [↓Results]

1 Y or RGB denotes the evaluation on Y (luminance) or RGB channels.

Stage 2 models for the NTIRE19 Competition

Download Models from Google Drive

Model name[Test Set] PSNR/SSIM
EDVR_REDS_SR_Stage2[REDS] (RGB) / [↓Results]
EDVR_REDS_SRblur_Stage2[REDS] (RGB) / [↓Results]
EDVR_REDS_deblur_Stage2[REDS] (RGB) / [↓Results]
EDVR_REDS_deblurcomp_Stage2[REDS] (RGB) / [↓Results]

DUF

The models are converted from the officially released models.
Download Models from Google Drive

Model name[Test Set] PSNR/SSIM1Official Results2
DUF_x4_52L_official3[Vid4] (Y4) 27.33/0.8319 [↓Results]
(RGB) 25.80/0.8138		(Y) 27.33/0.8318 [↓Results]
(RGB) 25.79/0.8136
DUF_x4_28L_official[Vid4]
DUF_x4_16L_official[Vid4]
DUF_x3_16L_official[Vid4]
DUF_x2_16L_official[Vid4]

1 We crop eight pixels near image boundary for DUF due to its severe boundary effects.
2 The official results are obtained by running the official codes and models.
3 Different from the official codes, where zero padding is used for border frames, we use new_info strategy.
4 Y or RGB denotes the evaluation on Y (luminance) or RGB channels.

TOF

The models are converted from the officially released models.
Download Models from Google Drive

Model name[Test Set] PSNR/SSIMOfficial Results1
TOF_official2[Vid4] (Y3) 25.86/0.7626 [↓Results]
(RGB) 24.38/0.7403		(Y) 25.89/0.7651 [↓Results]
(RGB) 24.41/0.7428

1 The official results are obtained by running the official codes and models. Note that TOFlow does not provide a strategy for border frame recovery and we simply use a replicate strategy for border frames.
2 The converted model has slightly different results, due to different implementation. And we use new_info strategy for border frames.
3 Y or RGB denotes the evaluation on Y (luminance) or RGB channels.