Video Super-Resolution Experiment

Source: /home/yifan/temp/vsr_exp/work on v100. Comparison of bicubic, residual-CNN L1, and CVVDP heatmap weighted residual-CNN variants.

Best JODres_cvvdp_hm_a1 · 7.820

Best PSNRres_cvvdp_hm_a1 · 26.127 dB

Lowest LPIPSres_cvvdp_hm_a1 · 0.5078

Metric Summary

Name	Kind	JOD	ΔJOD	PSNR	ΔPSNR	SSIM	L1	LPIPS	ΔLPIPS	Seconds
`res_cvvdp_hm_a1`	residual_cnn_cvvdp_hm_a1	7.820	+0.080	26.127	+0.093	0.640	0.03263	0.5078	-0.0139	57.7
`res_cvvdp_hm_a4`	residual_cnn_cvvdp_hm_a4	7.813	+0.073	26.071	+0.037	0.639	0.03293	0.5105	-0.0113	45.9
`res_l1`	residual_cnn_l1	7.792	+0.052	26.112	+0.079	0.636	0.03259	0.5127	-0.0090	43.1
`bicubic`	baseline_no_learning	7.740	+0.000	26.034	+0.000	0.628	0.03288	0.5217	+0.0000	-

Videos

Low-resolution input

Bicubic baseline

Residual CNN · L1

Residual CNN · CVVDP heatmap α=4

Residual CNN · CVVDP heatmap α=1

CVVDP heatmap on bicubic

Diagnostics

Frame comparison grid

Training loss curves

PSNR per frame