CVPR Daily - Tuesday

“ Typically, a network needs to convert this past information into an RGB color space representing video frames, which means if you want to interpolate 10 frames between the given two, you need to perform this refinement step 10 times , ” he explains. “ The total runtime is increased linearly along with that number. Our method simplifies the process, does the computation once, and can quickly interpolate any number of frames! ” This overall speed increase, particularly for high frame-rate interpolation, will benefit many practical applications in the real world, including video editing , video compression , and slow-motion animation . Being chosen for a poster at CVPR is no mean feat – what does Ping think made his paper stand out from the rest? “ In my opinion, there are three things: speed, accuracy, and design , ” he asserts. “ First, the speed and efficiency we’ve introduced, as I’ve described. We can interpolate one 720p frame at a speed of several milliseconds per frame. Second, it achieves good accuracy. In animation, it reaches close to state-of- the-art performance. Third, the design of the architecture is not very complicated, which means it can be easily applied to different scenarios. ” 9 DAILY CVPR Tuesday Ping Hu Examples of the MRN’s output (N = 4). S0 shows low reliability in areas with complex motion as intuitively expected. Our method refines initial flow with better details, and decomposes complex motion with shade changes (as indicated by the red circle) into multiple motion fields.