LinkNet_chainer

Implementation of LinkNet by chainer
Feature Forwarding: Exploiting Encoder Representations for Efficient Semantic Segmentation link

######## Training by cityscapes ########
# Calculate class balancing
python calculate_class_weight.py [mean or loss] --base_dir data_dir --result name --source ./pretrained_model/data.txt --num_classes 19 --dataset [cityscapes or camvid]
# Training encoder by cityscapes
python train.py experiments/enc_dec_paper.yml

######## Evaluate by cityscapes ########
python test.py experiments/test.yml

######## Visualize by cityscapes ########
python demo.py experiments/test.yml --img_path img.png

Implementation

• Spatial Dropout using cupy
• Baseline, model architecture
• Evaluate by citydataset
• Calculate class weights for training model
• Poly leraning rate policy

Requirement

• Python3
• Chainer3
• Cupy
• Chainercv
• OpenCV

TODO

• Load pretrained model of resnet18 (if necessary)