This is a repo contains tensorflow implementations of current state-of-art semantic segmentation nets, and expand them to dense pixel tasks including surface normal and depth estimation from monocular image.
Some of them are the latest Cvpr2017 receiving papers and some of them have their influence in the dense pixel tasks.
Mainly it contains following nets
Further More I will try to deploy the google MobileNets(MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications) to some of the above nets to see whether it can bring speed up without lossing much accuracy.
Currently only part of them are completed, as
- common this folder contains some common tool for all nets used to do the training job.
- enet contains the full implementation of the ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation in tensorflow.
- segnet contains the full implementation of the SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation in tensorflow.
Note that this repo uses PyYam config file to config the project. Make sure that it has been installed if you want to use the config file.
python train.py -c ./conf.yml
python deploy.py -i[input_dir] -o[output_dir] -m[model_file]
For more detail, refer to the code.
Note this project is released for personal and research use only. For a commercial license please contact the original paper authors.