/awareutils

A collection of useful utilities

Primary LanguagePythonMIT LicenseMIT

awareutils

A collection of Python utilities used within Aware Group. Generally computer vision for now. The objective is:

  • Reduce common errors or sources of confusion (see below).
  • Improve development time by a) getting features for free (e.g. threaded cameras) and b) getting 'debug' features for free (like printing FPS) which will (hopefully) help when issues are encountered.
  • Improve quality (performance etc.).
  • Standardize code across various projects.
  • Being opinionated and explicit. This helps deliver the above objectives, and speed up learning.

Installation

git clone https://github.com/awaregroup/awareutils
pip install ./awareutils[all]

The above installs everything including e.g. OpenCV and Pillow. If you know you're only going to work with OpenCV then this library supports that - in this case you'd pip install awareutils[cv2]. See extras-requirements.txt for a breakdown.

Features

  • Single Img class. The key features are:
    • Clarity between BGR / RGB / PIL, and no performance cost for using a particular one. If you only ever want PIL you don't need to install OpenCV.
    • Ability to save metadata (e.g. annotations) in the image itself (via EXIF metadata).
  • Various video classes which:
    • Are threaded for optimal performance (i.e. your main loop isn't hanging on decoding).
    • Can (hopefully) be easily extended for other cameras - just define how to open/close the camera, and read a frame.
    • OpenCV specific classes for:
      • Reading from a file - you can choose to read every frame sequentially, or simulate it as a live camera with a given FPS.
      • Reading from a live (USB) source, including nicely setting attributes like height etc.
      • Writing to a file. Eventually there'll be a nice way to get the 'right' fourcc without all the messing round.
  • One standard approach to coordinate systems. Note OpenCV vs numpy indexing or normalised vs not coordinates, etc.
  • Shapes ...
  • Nicer drawing functionality (and reasonably consistent between PIL and OpenCV) with real colours, multi-line and word-wrapped text, etc.
  • A nicer GUI experience than cv2.imshow with built in support for nice things like padding to full screen, a 'console' for informational text messages, etc.

You can see more about some of the design decisions in ./doc/decisions.md.

Bugs we're trying to cover

  • General bugs by rolling your own = ) Use this simply 'cos it's been used before, and is tested.
  • Usually cv2.rect(...) etc. edit the image in place ... except when it's not contiguous. Hence we use np.ascontiguousarray etc. when we're looking at arrays.
  • Confusiong around RGB vs BGR. So we're explicit.
  • Confusion between PIL and OpenCV. Support both, and conversion between.
  • Separation of annotations from image - so we allow saving metadata in the image EXIF metadata.
  • When reading an OpenCV camera, the buffer can lag - so do it in a thread. Likewise, thread up all the things for performance. This is surprisingly hard to do well and cover all cases.
  • On that note, OpenCV and threads can be lots of fun e.g. some operations have to stay in the same thread context etc. which can result in weird bugs (e.g. .geting attributes from a VideoCapture from another thread).
  • Pixel coordinates (vs normalised) are confusing. See above. Fix by being a little explicit and only using pixel coordinates, not using the term "Point", etc.
  • Invalid shape coordinates (e.g. outside image etc.). Check for this, and fix, etc.
  • Bounding box formats (x0y0x1y1, x0y0wh, xcycwh, etc.) - be explicit. Likewise (w, h) vs (h, w) - be explicit by forcing kwargs.
  • Annoyance setting up OpenCV windows fullscreen etc. Provide some utils.
  • Overlaying text on top of images with cv2.imshow can be annoying and often obscures the image etc. Our default GUI fixes this by having a 'console' for information, separate to the image.
  • Writing an image to a VideoWriter silently fails if the dimensions don't match. That's annoying to debug.

Developing

pip install -e .

Then you can edit at will, and ensure you run

pytest .

If you want to see some logs, ensure you do logger.enable('awareutils').