Download models if not existing

[Tobias-Fischer]

Allow running estimate_gaze.py etc without having to run download_models.py; in that case, download the models at runtime.

[ahmed-alhindawi]

In addition, we now have diverging download structures, one in CMakeLists.txt and the other in the download_models.py. I have toyed with the idea of executing download_models.py once catkin_make is run but the output of execute_process isn't inlined with the output of catkin_make making it look like the whole thing has frozen, when it hasn't....

If we do move towards download_models.py as being our only way of downloading model files, we should think about removing the catkin_download bit in CMakeLists.txt

[Tobias-Fischer]

Yes, I agree that we should remove the catkin_downloads altogether .. when I first did this there was only one model so it was fairly quick, but now it has exploded and we shouldn't do it during the build process anymore.

[Tobias-Fischer]

I thought about this again. I suggest the following:

• Split up download_models.py into methods download_landmark_models(), download_gaze_tensorflow_models(), download_gaze_pytorch_models() and download_blink_models(). The __main__ of download_models.py then simply calls all of these methods.
• In estimate_gaze.py and estimate_gaze_standalone.py, depending on the backend chosen, either call download_gaze_pytorch_models() or download_gaze_tensorflow_models().
• Similarly do the same thing for the blink and landmark nodes.
• Only check the md5sum in request_if_not_exist after downloading the model and not every time to avoid introducing start-up latencies.

What do you think @ahmed-alhindawi? If you're happy with the suggestion I'll go ahead with it.

[ahmed-alhindawi]

Sounds great!