Lossy GIF Compression Using Deep Intrinsic Parameterization

Anuj Pahuja, Simon Lucey

Abstract: With billions of animated GIFs being shared and viewed every day, it has become imperative for GIF hosting websites to serve content with minimal lag. To cater to the ever-decreasing attention span of a wide audience with different connectivity issues, it makes sense to suitably compress GIFs during transmission. We present a unique and interpretable approach to lossily compress GIFs or any temporal sequence of frames through a CNN based image parameterization technique and a simple scalar quantization scheme. Contrary to learned compression techniques, our approach is instance-specific and self-supervised.