- Repeat the FCN out of learning and researching purpose, especially for autonomous driving scenario.
- Due to the limitation of my GPU memory (2GB), I only implement a version of FCN with resNet34.
- I train the FCN with CityScape Dataset, the overall mean iou of fine label in CityScape Dataset is about 27%, the benchmark result with resNet101 is 30.4%, thus the performance is acceptable.
- It should be noted that the state-of-the-art performance of mean IoU is 84.5%
- matplotlib==3.1.3
- numpy==1.18.1
- torch==1.3.1
- torchvision==0.4.2
- Download the CityScape Dataset from here
- You will have to create an account with your email address to download the dataset
- Copy the dataset to
./data, the file sturcture of the workspace can be like the following:
- Setting the training parameter (like learning rate, weight decay, number of epoches, batch sizes, etc)
- Just run
train.py
- Just run
validation.py
- Then run
test.py - Two examples from RobotCar Dataset are provided in
./test
- Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2015: 3431-3440.