fwdekker / facemation Goto Github PK

Currently, Facemation supports only .jpg files. There's not really any good reason for this, since all operations by Facemation are over raw data read by libraries which don't care about the data type. The caches can remain JPG, but Facemation should support reading non-JPG inputs.

The README could be a lot nicer if there were animated examples to be shown to the user. Then the user knows what they can expect. A list of features would also be nice. And once all that has been added, a table of contents will probably be required.

When a user invokes Facemation from the CLI with options (e.g. ./facemation --help or ./facemation --faster (the latter doesn't do anything special)), the user should be shown text indicating that Facemation has no command-line options, and should tell the user how to change the configuration, and with a link to the config_default.py file as a reference for how the config works.

Using GitHub actions, I should be able to build for macOS as well, and once I have that set up, I can also use that to build for Windows and Linux, which will save me from booting VMs every time I want to do a release.

While face detection is done in parallel, normalisation and captioning are currently not. It should be relatively easy to make these parts parallel. However, the speedup may be minimal, since it's mostly about IO.

If input images were present before but no longer, cached items will still remain in the output/cache/ folder. Facemation should have an option so that all cache items that do not correspond to an image that is currently in the input is deleted.

I think the easiest way to add such a feature is by making a new section in the config, named general (or something), which is given to each stage of the pipeline.

Caches should be versioned, so that existing users can upgrade to a newer version of Facemation, and old cached data will not be used if it is incompatible with the newer version.

If photos are not taken regularly (say, every day for the first 6 months, and every 3 days for the last 6 months), then the speed of changes will be different in different parts. To counteract this, Facemation should (optionally) duplicate (or drop? (or both?)) frames so that the average rate of photos is the same everywhere.

If 0 or >=2 faces are detected, Facemation throws an exception and exits. This is frustrating if one has hundreds of images with many faulty ones. Instead, Facemation should continue running but collect the images, and then fail once the face detection stage has completed, and then output all errors at once.

If the input directory contains duplicate images, warn the user by indicating the corresponding files' names, as this may be an error. (Alternatively, give an error, and allow the user to ignore this error with a configurable setting.)