It would be interesting to see optimization of compression that's targeted at computer vision processing instead of human visual preferences. What I mean is tha

It would be interesting to see optimization of compression that's targeted at computer vision processing instead of human visual preferences. What I mean is that instead of making the image look as nice as possible at a certain file size, the goal would be to improve computer vision task performance scores, like running a classifier or segmentation model on that image and evaluating the correctness of the model prediction. A compression algo is better if the model accuracy is higher at a certain file size. Edit: yes, this has been done https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123650307.pdf https://www.ecva.net/papers/eccv_2020/papers_ECCV/papers/123... Jinyoung Choi and Bohyung Han, Task-Aware Quantization Network for JPEG Image Compression. ECCV 2020