Huge feature (200+ commits from different developers). It completely changes layout of data (please expect very long DB migration process if you have a lot of tasks). The primary idea is to send data as zip chunks (e.g. 36 images in one chunk) or encoded video chunks and decode them on the client side. It helps to solve the problem with latency when you try to view a separate frame in the UI quickly (play mode).
Another important feature of the patch is to provide access to the original images. Thus for annotations the client uses compressed chunks but if you want to export a dataset Datumaro will use original chunks (but video will be decoded with original quality and encoded with maximum/optimal quality in any case).
* Fix HTTP 400 error if together with vision data the user submit non-vision data (e.g. text files)
* Ignore SVG images because Pillow doesn't work with them.
Maya
Andrey Zhavoronkov
Thomas Albrecht
Nikita Manovich
Nikita Manovich
Ben Hoff