This repo contains the code for the short paper titled "Task-aware Image Compression with Autoencoder Networks". The basic idea of the paper was evaluating whether outputting a (binary) mask reconstruction along with the standard RGB reconstruction in autoencoder models would result in increased performance in a color detection task.
Model with single output: Raw (RGB) reconstruction
Model with two outputs: Raw (RGB) and Mask reconstruction
Table of results for both architectures at different image compression rates (bytes per frame):
Example of reconstructions. On the left side is the original RGB and Mask. On the right side is the reconstructed RGB and Mask:
Right now, the repo is bascially just a storage for the python scripts. Later it should be cleaned up and presented in a nicer way, with an example dataset and better description of the method.