amodal_road

1. amodal road segmentation using amodal datasets
2. code modified based on DANet

Usage

1. Install pytorch

   • The code is tested on python3.6 and torch 1.4.0.
   • The code is modified from DANet.

2. Clone the resposity

   git clone https://github.com/coolgrasshopper/amodal_road_segmentation.git 
   cd amodal_road_segmentation

3. Key files related to training and model construction:

• experiments/segmentation/encoding/nn/loss.py: define the cross entropy loss that is used for training
• experiments/segmentation/encoding/model/sseg/danet.py: define the proposed model architecture
• experiments/segmentation/encoding/nn/da_att.py: the Positional Attention module (PAM) and the Channel Attention Module
• experiments/segmentation/train.py: training code
• experiments/segmentation/test.py: testing code
• overlay.py: evaluation code of far and close IOU

4. training:

   cd experiments/segmentation
   CUDA_VISIBLE_DEVICES=0,1 python3 train.py --model danet --backbone resnet50 --checkname danet50 --base-size 1024 --crop-size 768 --epochs 240 --batch-size 8 --lr 0.003 --workers 16  

5. testing:

   cd experiments/segmentation
   CUDA_VISIBLE_DEVICES=0 python3 test.py --model danet --backbone resnet50 --base-size 1024 --crop-size 768 --batch-size 8 --workers 16  

TODO:

1. update evaluation code to the repo

acknowlegement

semantic-foreground-inpainting

DANet