A Collection of Algorithms and Datasets for Stereo Image Super-Resolution
With recent advances in stereo vision, dual cameras are commonly adopted in mobile phones and autonomous vehicles. Using the complementary information provided by binocular systems, the resolution of image pairs can be enhanced by stereo image super-resolution (SR) algorithms. In this repository, we first present a collection of datasets and papers on stereo image SR, together with their codes or repos. Then, we develop a benchmark to comprehensively evaluate milestone and state-of-the-art methods. Welcome to raise issues regarding our survey and submit novel results (better together with original files and source codes) to our benchmark.
Note: This repository will be updated on a regular basis, so stay tuned~~🎉🎉🎉
News:
2022-04-20: We add a new work Trans-SVSR that is accepted to CVPRW 2022.
We benchmark several methods on the KITTI 2012, KITTI 2015, Middlebury and Flickr1024 datasets. The test sets used below can be downloaded from Google Drive and Baidu Drive (Key: NUDT). Note that, All these methods have been retrained on the same training set (60 images from the Middlebury dataset and 800 images from the Flickr1024 dataset) for fair comparison.
PSNR and SSIM metrics are used for quantitative evaluation, which are first calculated on each view (without boundary cropping) independently, then averaged between left and right views to generate the score of a scene. Finally, the score of a dataset is obtained by averaging the scores of all its scenes.
PSNR and SSIM values achieved by different methods for 2xSR:
Method
Scale
#Params.
KITTI 2012
KITTI 2015
Middlebury
Flickr1024
Bicubic
2×
—
28.51/0.8842
28.61/0.8973
30.60/0.8990
24.94/0.8186
VDSR
2×
0.66M
30.30/0.9089
29.78/0.9150
32.77/0.9102
25.60/0.8534
EDSR
2×
38.6M
30.96/0.9228
30.73/0.9335
34.95/0.9492
28.66/0.9087
RDN
2×
22.0M
30.94/0.9227
30.70/0.9330
34.94/0.9491
28.64/0.9084
RCAN
2×
15.3M
31.02/0.9232
30.77/0.9336
34.90/0.9486
28.63/0.9082
StereoSR
2×
1.08M
29.51/0.9073
29.33/0.9168
33.23/0.9348
25.96/0.8599
PASSRnet
2×
1.37M
30.81/0.9190
30.60/0.9300
34.23/0.9422
28.38/0.9038
BSSRnet
2×
1.89M
31.03/0.9241
30.74/0.9344
34.74/0.9475
28.53/0.9090
iPASSR
2×
1.37M
31.11/0.9240
30.81/0.9340
34.51/0.9454
28.60/0.9097
SSRDE-FNet
2×
2.10M
31.23/0.9254
30.90/0.9352
35.09/0.9511
28.85/0.9132
NAFSSR-T
2×
0.45M
31.26/0.9254
30.99/0.9355
35.01/0.9495
28.94/0.9128
NAFSSR-S
2×
1.54M
31.38/0.9266
31.08/0.9367
35.30/0.9514
29.19/0.9160
NAFSSR-B
2×
6.77M
31.55/0.9283
31.22/0.9380
35.68/0.9544
29.54/0.9204
NAFSSR-L
2×
23.8M
31.60/0.9291
31.25/0.9386
35.88/0.9557
29.68/0.9221
PSNR and SSIM values achieved by different methods for 4xSR:
Method
Scale
#Params.
KITTI 2012
KITTI 2015
Middlebury
Flickr1024
Bicubic
4×
—
24.58/0.7372
24.38/0.7340
26.40/0.7572
21.82/0.6293
VDSR
4×
0.66M
25.60/0.7722
25.32/0.7703
27.69/0.7941
22.46/0.6718
EDSR
4×
38.9M
26.35/0.8015
26.04/0.8039
29.23/0.8397
23.46/0.7285
RDN
4×
22.0M
26.32/0.8014
26.04/0.8043
29.27/0.8404
23.47/0.7295
RCAN
4×
15.4M
26.44/0.8029
26.22/0.8068
29.30/0.8397
23.48/0.7286
StereoSR
4×
1.08M
24.53/0.7555
24.21/0.7511
27.64/0.8022
21.70/0.6460
PASSRnet
4×
1.42M
26.34/0.7981
26.08/0.8002
28.72/0.8236
23.31/0.7195
SRRes+SAM
4×
1.73M
26.44/0.8018
26.22/0.8054
28.83/0.8290
23.27/0.7233
BSSRnet
4×
1.91M
26.47/0.8049
26.17/0.8075
29.08/0.8362
23.40/0.7289
iPASSR
4×
1.42M
26.56/0.8053
26.32/0.8084
29.16/0.8367
23.44/0.7287
SSRDE-FNet
4×
2.24M
26.70/0.8082
26.43/0.8118
29.38/0.8411
23.59/0.7352
NAFSSR-T
4×
0.46M
26.79/0.8105
26.62/0.8159
29.32/0.8409
23.69/0.7384
NAFSSR-S
4×
1.56M
26.93/0.8145
26.76/0.8203
29.72/0.8490
23.88/0.7468
NAFSSR-B
4×
6.80M
27.08/0.8181
26.91/0.8245
30.04/0.8568
24.07/0.7551
NAFSSR-L
4×
23.8M
27.12/0.8194
26.96/0.8257
30.20/0.8605
24.17/0.7589
Recources
We provide some original super-resolved images and useful resources to facilitate researchers to reproduce the above results.
The 2x/4x models of EDSR/RDN/RCAN retrained on stereo image datasets. Google Drive, Baidu Drive (Key: NUDT).