VoVNet.DeepLabV3

This is a pytorch implementation of DeepLabV3 with VoVNet Backbone Networks. This code based on pytorch implementation of DeepLabV3.pytorch.

Highlights

• Memory efficient
• Better performance
• Faster speed

Comparison with ResNet & DenseNet backbones

• For fair comparison, totally SAME training setup except for backbone
• 50 epoch
• 0.007 base_lr
• 16 batch size
• same ASSP module & parameters
• V100 GPU
• pytorch 1.1.0a0+3752916
• CUDA v10
• cuDnn v7.3

Backbone	mIoU	inference time (ms)	Memory usage (MB)	Energy Efficiency (J/frame)	DOWNLOAD
ResNet-50	74.27	24	2193	4.1	link
DenseNet-201	75.63	50	3945	7	link
VoV-39	75.71	19	1901	3.1	link
ResNet-101	76.81	32	2865	15.8	link
DenseNet-161	76.13	49	4523	8.3	link
VoV-57	77.4	25	2251	4.2	link

ImageNet pretrained weight

• VoVNet-39
• VoVNet-57

Preparation

git clone https://github.com/stigma0617/VoVNet-DeepLabV3.git
cd VoVNet-DeepLabV3

mkdir -p data/pretrained
cd data/pretrained
wget https://www.dropbox.com/s/b826phjle6kbamu/vovnet57_statedict_norm.pth
wget https://www.dropbox.com/s/s7f4vyfybyc9qpr/vovnet39_statedict_norm.pth

PASCAL VOC 2012 Dataset

cd ~/VoVNet-DeeplabV3/data
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
tar -xf VOCtrainval_11-May-2012.tar
cd VOCdevkit/VOC2012/
wget http://cs.jhu.edu/~cxliu/data/SegmentationClassAug.zip
wget http://cs.jhu.edu/~cxliu/data/SegmentationClassAug_Visualization.zip
wget http://cs.jhu.edu/~cxliu/data/list.zip
unzip SegmentationClassAug.zip
unzip SegmentationClassAug_Visualization.zip
unzip list.zip

Training

Specifying a backbone network with --backbone,

For VoVNet-39, --backbone vovnet39

python main.py --train --exp bn_lr7e-3 --epochs 50 --base_lr 0.007 --backbone vovnet39

Evaluation

use the same command except delete --train

wget https://www.dropbox.com/s/oqqozntgrowmfb1/deeplab_vovnet39_pascal_v3_bn_lr7e-3_epoch50.pth -P data/
python main.py --exp bn_lr7e-3 --epochs 50 --base_lr 0.007 --backbone vovnet39