Implementation of ICNet by Tensorflow
ICNet is a compressed-PSPNet Based image cascade network, proposed by H. Zhao et al. ICNet for Real-Time Semantic Segmentation on High-Resolution Images. Paper is here . This network is characterized by its runtime and multi-scale input architecture.
Accuracy is measured by Cityscapes data.
This network runs at the speed of 30fps with single GPU. The network is largely composed of simple 3 layers encoding network and 4 layers decoding network.
Different scale of layers are fused in CFF(Cascade Feature Fusion) box.
Dilated convolution encourages a network to dense labeling and enlarge the receptive field.
In tensorflow, it implemented by using tf.nn.astrous
This picture helps understand how dilate convolution works. Depending on dilation value, the filter for convolution become sparse. As a result, it leads to expand receptive fields of layers.
Trained with 100 road images.
- Python 3.5.2
- Tensorflow 1.3.0
- Numpy 1.13.1
- Scipy 0.19.1
At this directory.
mkdir data
mkdir output
Example data is from KITTI road dataset.
Download from here(80Mb).
Download road folder.
Need to put road folder under ./data folder.
1, Make sure you have training data under ./data folder.
2, python main.py
You will have an output example in ./output/post_train_output.png .
Network architecture should be really similar to the paper. I was supposed to train whole data from KITTI, but it was hard to train whole data due to the lack of hardware resource. I submit codes for training a hundred images.