How to speed up predictions about kittiseg

[MarvinTeichmann]

The inference time is 24 images per second. This time is archived if the model is build once and images are feed one by one via feed_dict. This is done in eval.py. Don't use demo.py for large inference. Also see the discussion here.

For an even greater speed-up you could use queues and batch processing of inference images. Eval.py is written to simulate a real-time system which does not come with batch-processing.

[jbrozena22]

Thanks so much, that's very helpful. Also more generally, thank you for this amazing project.