Project Description

This project enhances video frame interpolation by using a convolutional neural network (CNN) to estimate two 1D kernels, reducing memory requirements compared to traditional 2D kernel methods. We extracted consecutive frames from YouTube videos and applied data augmentation techniques to create a training dataset.

A custom loss function combining L1 loss, VGG19 features, and SSIM was developed to improve perceptual accuracy, resulting in sharper interpolated images. The model performed well on real-life videos with moderate motion, though challenges arose with larger motions and high-resolution inputs. This work provides an efficient solution for high-quality video frame interpolation.