EL-GAN

This repository offers an implementation for the Embedded loss GAN. The network was introduced in the paper EL-GAN: Embedding Loss Driven Generative Adversarial Networks for Lane Detection by Mohsen Ghafoorian, Cedric Nugteren, Nora Baka, Olaf Booij, Michael Hofmann.

This was part of my FYP for applying to the Computer and Communications Engineering degree at the Lebanese University, Faculty of Engineering.

How to use

After downloading the code from this repository you can download the third party dependencies from requirements.txt. Then you have the model implmentation as well as a training code that you can edit to fit your needs. The dataset class could also be easily adapted for other tasks.

Introduction

The model introduced relied heavily on the Denseblocks which were used in both the generator and the discriminator, it also introduced a novel approach to the discriminator, where the loss was caculated as a distance between the embeddings of the fake labels (labels produced by the generator), and the real labels (ground truth). They called this loss the embedding loss, they also passed not only the labels to the discriminator, but also the original images the labels respond to, creating a discriminator with "two heads". Note: Labels in the above paragraph and later on refers to the 2-channel image containing the lane markings.

Generator

The generator in the model was a FC-DenseNet (fully connected dense network), which was introduced in this paper. The architecture of the generator could be seen in the image below.

The downwards path is created by connecting Down_n blocks and Up_n blocks, a donwn (up) block is a dense block with n denselayers and a tranistion down (up) blocks. There architecure can be found in the original repository by Fredrico Baldassarre. As seen in the image above the layers in the down path are [1, 2, 3, 4, 5, 6, 8] layers blocks, there an 8 layers block in the middle, and [8, 6, 5, 4, 3, 2, 1] layers blocks in the up path.

Discriminator

The discriminator in EL-GAN (as mentioned before) has two heads, they take in the original images and the labels

The two head converge after 2 dense blocks, where there output is concatenated and continues through 5 more dense blocks. The embedding is the output of the last dense block (the 1x1 conv and the classification layer are there more for symmetry than anything else, they aren't necessary in the network.

For later

The repository still needs some work, so I may return to it later.

• The training code needs to be updated to what I used for training (I trained on a colab notebook).
• The parameters of the model should be change to better resembe what was used in the original paper.
• I may add some documentation to the functions and classes to make it easier to understand and work with.
• A complete training is yet to be implemented, I didn't do it because I lacked the hardware resources necessary.
• I may try the model for other semantic segmentation tasks.
• If anyone have any suggestions, or improvements, please feel free to let me know.

Thanks

The repository was built on top of this repository owned by Federico Baldassarre - the referenced repository offered an implementation for FCDenseNets and DenseNets. I would like to thank the author, whose repository not only helped me for the implementation and getting a better understanding for DenseNets, but also with organizing and presenting a project on GitHub.
I would also like to thank the author's of EL-GAN, especially Mr. Ghafoorian for answering an enquiry I had on email, in addition to the authors of "The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation" and "Densely Connected Convolutional Networks", these people are innovating in research, and by sharing their work, helping people like me to get to know and work in the field (even if I barely scratch the surface).
The images for the architecture were created using app.diagrams.net.