Wow, using semantic interpretation of an image as a lossy compression technique. That’s cool. Our own memories probably work something like this. Compressing a 25 megapixel HDR image to “dog” is a compression ratio of over 300 million, but is pretty lossy. 🙂
Originally shared by Vincent Vanhoucke
When I joined Stanford’s Compression and Classification Group in 1999, it became quickly evident to me that research in signal compression was really at an impasse: it was clear at the time that one would have to move towards more semantic interpretations of images and videos to make any significant gains in bandwidth, and in spite of standards already moving towards enabling these ‘higher-level’ coding methods, nobody really knew how to go about them.
Fast forward to today, I’m very excited to see deep nets make a significant dent into the problem, while enabling seamless, practical variable-rate coding, bit-per-bit progressive decoding, and with huge gains over JPEG to boot.