Computer Vision News - January 2022

26 Congrats, Doctor! Fabian Mentzer defended his PhD in January 2021 at ETH Zurich and was awarded the “ETH medal for outstanding doctoral theses” for his work. He now works as a Research scientist at Google, focusing on neural compression. This is also the topic he worked on during his PhD, his thesis was on lossy and lossless image compression with neural networks. Congrats, Doctor Fabian! When (lossily) compressing images, the goal is to store an input image in a file so that on one hand, the file is small, and on the other hand that when we open the file , the reconstructed image looks as close as possible to the input image. There is a trade-off here: we can always transmit more information, thereby making the file larger, and the reconstruction becomes better. Inversely, bits can be saved by ignoring part of the image – something everyone that has ever seen a blocky JPEG image on the internet is familiar with. To solve this problem with neural networks , there are a few challenges, one interesting one is described next. On a high level, we want to formalize the above trade-off, and we want to tell the optimization algorithm which parts of the image can safely be ignored to save bits. Intuitively, we are looking for a loss that is aware of how humans perceive images : Humans do not like blocky images and can tell that those are compression artifacts very easily. Same goes for blurry images. However, if the changes that happen due to compression are more semantic, they are hard to notice. For example, when