Your child’s first steps, your daughters dancing performance or your son’s fantastic skateboard tricks. These are the moments you will never forget and still, the pass in the blink of an eye. Do you ever wish you would have captured these moments not only on video but in slow motion to enjoy every moment even more?

Nowadays, videos with 240fps (frame-per-second) can be recorded from a mobile phone, while videos with higher frame rates still require a high-speed camera. Also, these high frame rates need an enormous storage capacity and are not feasible in a small mobile device due to their high energy consumption. So how can you ensure to capture and relive these incredible moments often in slow motion – unpredictable and unplanned – without carrying around a sizeable high-speed camera all the time?

Scientists at the University of Massachusetts at Amherst, NVIDIA and the University of California have just found the perfect solution. They successfully conducted a research project in which they used Artificial Intelligence to convert videos with standard frame rates to super slow motion videos. To achieve this, they propose an end-to-end CNN (Convolutional Neural Network) model for frame interpolation that has an explicit sub-network for motion estimation. At first, they used a flow computation CNN to estimate the bidirectional optical flow between the two input frames. In a next step, then two flow fields are linearly fused to refine the approximated flow fields and predict soft visibility maps for interpolation.

Their scientific research has been recently published with the title “Super SloMo: High-Quality Estimation of Multiple Intermediate Frames for Video Interpolation” where they introduce all the necessary steps and experimental results.

Last but not least, speaking about videos, to gain a broader idea about this exceptional achievement, we would recommend you to watch the highly informative video that was recently presented at the renowned CVPR 2018: Computer Vision and Pattern Recognition.

Source: https://arxiv.org/pdf/1712.00080v1.pdf