Semantic Segmentation using FPN

This repository is a semantic segmentation part implement of Kaiming He, Panoptic Feature Pyramid Networks .

To do

• Semantic Segmentation Branch
  • Multiple GPUs training
  • Train on CamVid Dataset
  • Train on Cityscapes Dataset
  • Train on NYUD v2 Dataset
  • Train on PASCAL Context Dataset

results

Not as good as the result in the paper, I am tring to improve it.

Dataset	mIoU	Pixel Acc	FWIoU	Backbone	Trained model
CamVid	0.570	0.920	0.861	ResNet101	CamVid
Cityscapes	0.605	0.928	0.872	ResNet101	CityScapes

Training

Prepare data

• default dataset is CamVid

download pytorch 1.0.0 from pytorch.org

download CamVid dataset or Cityscapes dataset

• for CamVid dataset, make directory "data\CamVid" and put "701_StillsRaw_full", "LabeledApproved_full" in "CamVid", then run:

python data/CamVid_utils.py    

• for Cityscapes dataset, make directory "Cityscapes" and put "gtFine" in "Cityscapes/gtFine_trainvaltest" folder, put "test", "train", "val" in "Cityscapes/leftImg8bit" foloder.

Train the network

train with CamVid dataset:

change to your own CamVid dataset path in mypath.py, then run:

python train_val.py --dataset CamVid --save_dir /path/to/run

for multiple GPUs training, change to your own CamVid dataset path in mypath.py, then run:

python train_val.py --dataset CamVid --save_dir /path/to/run --mGPUs True --gpu_ids 0,1,2

train with Cityscapes(default) dataset: change to your own CityScapes dataset path in mypath.py, then run:

python train_val.py --dataset Cityscapes --save_dir /path/to/run

for multiple GPUs training, change to your own CityScapes dataset path in mypath.py, then run:

python train_val.py --dataset Cityscapes --save_dir /path/to/run --mGPUs True --gpu_ids 0,1,2

Test

Test with CamVid dataset(val), run:

python test --dataset CamVid --exp_dir /path/to/experiment_x

Test with Cityscapes dataset(val), run:

python test.py --dataset Cityscapes --exp_dir /path/to/experiment_x

If you want to plot the color semantic segmentation prediction of the test input color image, please set --plot=True, for example:

python test.py --dataset Cityscapes --exp_dir /path/to/experiment_x --plot True

Acknowledgment

FCN-pytorch

pytorch-deeplab-xception

pytorch-fpn

fpn.pytorch