C3 and SINet for Lightweight segmentaiton model on Cityscape dataset

Please code to this link for full code

Requirements

• python 3.6
• pytorch >= 0.4.1
• torchvision>=0.2.1
• opencv-python>=3.4.2.17
• numpy
• tensorflow>=1.13.0
• visdom

Model

Hyojin Park, Youngjoon Yoo, Geonseok Seo, Dongyoon Han, Sangdoo Yun, Nojun Kwak " C3: Concentrated-Comprehensive Convolution and its application to semantic segmentation " (paper)

Hyojin Park, Lars Lowe Sjösund, YoungJoon Yoo, Nicolas Monet, Jihwan Bang, Nojun Kwak " SINet: Extreme Lightweight Portrait Segmentation Networks with Spatial Squeeze Modules and Information Blocking Decoder" (paper)

Model	# of Param(M)	# of Flop(G)	size for Flop	IoU( val )	IoU (test)	server link
C3Net[2,3,7,13]	0.19	3.15	512*1024	66.87	64.78	link
C3NetV2[2,4,8,16]	0.18	2.66	512*1024	66.28	65.48	link
SINet	0.12	1.22	1024*2048	68.22	66.46	link

• C3NetV2 has same encoder structure with C3Net, but uses bilinear upsampling for a decodder structure.
• SINet is accepted in WACV2020.

Train

Once you train the model, my code automatically export format for Cityscape Testserver from best training model. I used P-40 GPU for training. C3 and C3_V2 require 2 GPU and SINet needs 1 GPU. Train validation txt is for datalodaer function here

python main_multiscale.py -c C3.json

python main_multiscale.py -c C3_V2.json

python main_Auxloss.py -c SINet.json

Citation

If our works is useful to you, please add two papers.

@article{park2018concentrated,
  title={Concentrated-Comprehensive Convolutions for lightweight semantic segmentation},
  author={Park, Hyojin and Yoo, Youngjoon and Seo, Geonseok and Han, Dongyoon and Yun, Sangdoo and Kwak, Nojun},
  journal={arXiv preprint arXiv:1812.04920},
  year={2018}
}


@article{park2019sinet,
  title={SINet: Extreme Lightweight Portrait Segmentation Networks with Spatial Squeeze Modules and Information Blocking Decoder},
  author={Park, Hyojin and Sj{\"o}sund, Lars Lowe and Monet, Nicolas and Yoo, YoungJoon and Kwak, Nojun},
  journal={arXiv preprint arXiv:1911.09099},
  year={2019}
}

Acknowledge

We are grateful to Clova AI, NAVER with valuable discussions.

I also appreciate my co-authors YoungJoon Yoo, Dongyoon Han, Sangdoo Yun and Lars Lowe Sjösund from Clova AI, NAVER, Nicolas Monet from NAVER LABS Europe and Jihwan Bang from Search Solutions, Inc

I refer ESPNet code for constructing my experiments and also appreciate Sachin Mehta for valuable comments. Sachin Mehta is ESPNet and ESPNetV2 author.