Skip to content
Project for Deep Learning course at École Polytechnique: creating a algorithm to do semantic segmentation on images and perform neural style transfer on user chosen segments
Jupyter Notebook Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
examples
semantic_segmentation
.DS_Store
LICENSE
README.md
Reference.docx
UseCasesNotebook.ipynb
final_deep_photo.pdf

README.md

DeepPhoto Project

This project is part of the course Deep Learning - MSc Data Science for Business - HEC Paris and Ecole Polytechnique

Team members

Ching-yu Lin

Fernado Perez

Jorgen Lund

Jiahao Wang

Roberta Conard

Usecase

Design a mobile application prototype that allows users to do fine-grained image editing in real-time before uploading the social applications on their mobile phones.

Techniques

We adapted the semantic segmentation technique (RefineNet) to identify the human part in images. Then we use the neural style transfer to transfer the part of image that user selected. For more details, please see the slides.

Final Result

We finetuned and pruned the RefineNet to our use case.

Our smallest model: LightNet-MobileNet acheived 0.809 mIoU with only 13MB size, which outperforms the off-the-shelf DeepLabV3 (0.806 mIoU, 233MB) in this specific task in terms of the quality, size and inference speed.

First five images in testset:

Full pipeline examples:

Duration

This project was conducted between 2019-Nov-05 and 2019-Nov-11.

Reference

The RefineNet model architecture is adapted from https://github.com/ansleliu/LightNet

You can’t perform that action at this time.