ENet_chainer

Implementation of ENet by chainer
ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation [link]

Result (Cityscapes)

Implementation	Image Size	Global accuracy	Class accuracy	mean IoU
Chainer(val)	512✕1024	92.59 %	71.49 %	59.1 %
Original(test)	1024✕2048	-----	----	58.3 %

Visualization

######## 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
・Single GPU
python train.py experiments/enc_paper.yml
・Multi GPUs
python train.py experiments/enc_paper.multi.yml

# Training decoder by cityscapes
python train.py experiments/enc_dec_paper.yml

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

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

Implementation

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

Requirement

• Python3
• Chainer3
• Cupy
• Chainercv
• OpenCV

TODO

• Convert caffemodel to chainer's model format
• Create merge function between convolution and batch normalization