- All dependencies are indicated in
env.txt
- Download weights of
Resnet-18
and put intocode/
- Link to weights of
Resnet-18
https://d2j0dndfm35trm.cloudfront.net/resnet-18.t7 The weights will be automatically loaded into respective layers in the beginning of training
- This example runs on cityscapes dataset, and data are prepared used official scripts. Link to cityscapes dataset https://www.cityscapes-dataset.com/
- Put your data into
data/
- All scripts are in
code/
The label mapping is inlabels.py
Runpython segmentation.py
to train a network, 5 latest checkpoint will be kept during train
If you prefer using Jupyter Notebook, you may look intosegmentation.ipynb
- LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation https://arxiv.org/abs/1707.03718
- Pytorch implimentation from the author https://github.com/e-lab/pytorch-linknet
- Torch implementation of the author (in Lua) https://github.com/e-lab/LinkNet
- Enet in Tensorflow https://github.com/kwotsin/TensorFlow-ENet