ICCV Daily 2021 - Tuesday

As far as Yinxiao knows, this is the first paper to look at compression-aware video super resolution in recent years. There are a few standard video testing datasets, but none of them have a heavy compression. Compressed videos are very challenging as depending on compression settings or parameters, high frequency information is smoothed during the compression and the degradation kernel is changed . The team tested their method thoroughly on real video use cases with a wide range of compression rates. They downloaded videos from YouTube , as they felt those everyday videos that we all watch are the kind that they should be targeting. The algorithms work well at improving the quality of such videos compared to other models. They did an ablation study comparing against the six most recent models and this work gave the best results. “ The design of our algorithms focuses on a few aspects, ” Yinxiao explains. “ We are fully aware of the current video compression strategy. First, we do a forward and backward pass to eliminate the error accumulation. Second, we do a detailed flow estimation because flow is very important in the video to use the temporal information to boost each of the frames. Third, we have a Laplacian enhancement module so that we can maximally preserve the high-frequency details in the video restoration. ” All three modules boost the video super-resolution results on the compressed videos. Those compressed videos are not visually distinguishable. If there are compression artifacts, other models amplify those artifacts, but this model removes them and upscale the frames . The team use TensorFlow for training the framework, which makes their work easily reproducible. They want others to explore it and use it to test their own videos. 9 DAILY ICCV Tuesday Yinxiao Li