tf-semantic-segmentation-FCN-VGG16
tf-semantic-segmentation-FCN-VGG16 copied to clipboard
Semantic segmentation for classifying road. "Fully Convolutional Networks for Semantic Segmentation (2015)" implemented using TF
Semantic Segmentation
- This repository is for udacity self-driving car nanodegree project -
Semantic Segmentation
. - Implement this paper: "Fully Convolutional Networks for Semantic Segmentation (2015)"
- See
FCN-VGG16.ipynb
Implementation Details
Network
FCN-8s
with VGG16
as below figure.
Dataset
- Kitti Road dataset from here.
Hyperparameters
Learning rate, batch size and keep probability were tunned by random search. If you want to see code for this: Link
- Optimizer:
Adam
- Learning rate:
0.0002395
- Deconvolution
l2 regularization
factor:1e-3
- Batch size:
2
- Training epochs:
30
-
Keep prob
for dropout (VGG):0.495
Results
Loss
After 30 epochs, loss became about 0.05
Nice results
These are pretty nice results. It seems like the network classify road area well.
Bad results
These are bad results. I believe that the results will be better using the following methods.
- Use more deeper network (e.g. ResNet)
- Augment given data or train network with another data (e.g. CityScape)
- Use different architecture (e.g. U-Net)
- Use post processing (e.g. CRF(Conditional Random Field))
Setup
Frameworks and Packages
Make sure you have the following is installed: