Support video import

[zhiltsov-max]

Visual datasets are not restricted by images and often come from video frames. The ability to use video data is essential for such project as Datumaro.

Requirements:

• Ability to import a video as a data source
  • A way to split video into frames
  • A way to utilize a video as a single dataset entry (action recognition tasks etc.)
• More natural support of track annotations and other video-specific types of annotations
• Ability to export a dataset as a video

[zhiltsov-max]

v1 design:

• For clients, who interested in just video reading - provide video reading options
  • CLI / API levels: add video frames extractor, should work like other import-only formats
  • API level: add video reader, a class to read video by frames
  • Use opencv as the backend (would use av, but it has problems with GPL precompiled binaries)
• For clients, who need stable results, which are independent of library versions - provide video splitting options
  • An instruction in docs and/or a command
  • Formats that use video frames either work with frame directories or provide such option.
• Issue a warning on addition of the video frames extractor in project. Recommend manual splitting for having stable results
• API level: DatasetItem still have either a video or a frame linked. No options to use both.
• Videos are not exported by recoding. When supposed by a format, only file copying is done.