This architecture consists of a ResNet followed by an Autoencoder. The goal is to compress both the image and its segmentation and transmit the resulting embeddings in the latent space through a mobile network. Additionally, there is a parameter that allows for selecting the size of the embeddings, which simulates a bottleneck in the network
conda create --name autoenc --file requirements.txt
python3 example.py
- Download the dataset from here (free registration needed). I reccomend to use
leftImg8bit_trainvaltest.zip
gtFine_trainvaltest.zip
. - Run
python3 train.py