NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study (supplementary material)

This is a supplement to our paper, “NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study” [1]. We report results for the top challenge methods [10]: SNU CVLab1 [6], SNU CVLab2 [6], Lab402 [2] and HelloSR [10] in comparison with VDSR [5] and A+ [12]. We use the classic setup (bicubic downscaling) of the challenge and report PSNR, SSIM [9], IFC [8], and CORNIA [13] results computed on Y channel from YCbCr or directly on the RGB super-resolved image for ×2, ×3, ×4 magnification factors. In addition to our newly introduced DIV2K dataset [1] we report also results for the most commonly used datasets in the recent literature: Set5 [3, 11], Set14 [14], B100 [7, 12], and Urban100 [4]. 1. DIV2K Test 100 images: quantitative results In Tables 1,2,and 3 we report average quantitative results over DIV2K test 100 images for magnification factors ×2, ×3, and ×4, respectively. Bicubic interpolation result is included for reference. Method computed on Y from YCbCr computed on RGB PSNR SSIM IFC CORNIA PSNR SSIM IFC CORNIA SNU CVLab1 36.38 0.953 8.11 13.4 34.93 0.948 8.22 13.0 SNU CVLab2 36.28 0.953 8.06 14.7 34.83 0.947 8.17 14.4 Lab402 36.12 0.952 7.97 13.6 34.66 0.946 8.07 13.3 HelloSR 35.87 0.950 7.91 15.1 34.43 0.944 8.03 14.7 VDSR[5] 35.09 0.943 7.28 17.7 33.47 0.935 7.36 17.4 A+[12] 34.26 0.935 7.98 29.5 32.64 0.925 7.15 29.4 Bicubic 32.44 0.910 6.31 45.9 30.98 0.899 5.76 45.6 Table 1. Quantitative results on the test set of DIV2K with ×2. Method computed on Y from YCbCr computed on RGB PSNR SSIM IFC CORNIA PSNR SSIM IFC CORNIA SNU CVLab1 32.59 0.900 4.97 21.2 31.13 0.889 5.01 20.9 SNU CVLab2 32.49 0.899 4.93 22.2 31.04 0.888 4.97 21.9 Lab402 32.29 0.896 4.80 22.8 30.83 0.884 4.83 22.5 HelloSR 32.20 0.895 4.78 23.4 30.74 0.883 4.82 23.1 VDSR[5] 31.48 0.881 4.40 27.0 29.89 0.866 4.43 26.9 A+[12] 30.89 0.869 4.61 40.3 29.32 0.852 4.26 40.3 Bicubic 29.66 0.837 3.67 58.3 28.20 0.821 3.43 57.9 Table 2. Quantitative results on the test set of DIV2K with ×3.