Gated-SCNN

A PyTorch implementation of Gated-SCNN based on ICCV 2019 paper Gated-SCNN: Gated Shape CNNs for Semantic Segmentation.

Requirements

• Anaconda
• PyTorch

conda install pytorch torchvision cudatoolkit=11.0 -c pytorch

• timm

pip install timm

• opencv

pip install opencv-python

• cityscapesScripts

pip install cityscapesscripts

Usage

Train model

python train.py --epochs 175 --backbone_type resnet101
optional arguments:
--data_path                   Data path for cityscapes dataset [default value is 'data']
--backbone_type               Backbone type [default value is 'resnet50'](choices=['resnet50', 'resnet101'])
--crop_h                      Crop height for training images [default value is 512]
--crop_w                      Crop width for training images [default value is 512]
--batch_size                  Number of data for each batch to train [default value is 4]
--epochs                      Number of sweeps over the dataset to train [default value is 60]
--save_path                   Save path for results [default value is 'results']

Eval model

python viewer.py --model_weight resnet101_800_800_model.pth
optional arguments:
--data_path                   Data path for cityscapes dataset [default value is 'data']
--model_weight                Pretrained model weight [default value is 'results/resnet50_512_512_model.pth']
--input_pic                   Path to the input picture [default value is 'test/berlin/berlin_000000_000019_leftImg8bit.png']

Results

The experiment is conducted on one NVIDIA TITAN RTX (24G) GPU, and there are some difference between this implementation and official implementation:

1. res2/res3/res4 are used in GCLs;
2. The non-differentiable part of dual task loss is not implemented;
3. The model is trained for 60 epochs with batch_size 4 on 512x512 images.

The left is the input image, the middle is ground truth segmentation, and the right is model's predicted segmentation.

ResNet50