Machine Learning-based Image Compression: Ready for Prime Time?

Computer Vision is undergoing dramatic changes because deep learning techniques are now able to solve complex non-linear problems. Computer vision pipelines used to consist of hand engineered stages mathematically optimized for some carefully chosen objective function. These pipelines are being replaced with machine-learned stages or end-to-end learning techniques where enough ground truth data is available. Similarly, for decades image compression has relied on hand crafted algorithm pipelines, but recent efforts using deep learning are reporting higher image quality than that provided by conventional techniques. Is it time to replaced discrete cosine transforms with machine learned compression techniques. This talk will examine practical aspects of deep learned image compression systems as compared with traditional approaches. We consider memory, computation and other aspects in addition to rate-distortion to see when ML-based compression should be considered or avoided. We also discuss approaches using a combination of machine learned and traditional techniques?