README.md

1.MSA-Net

Code for paper ‘Multi-unit stacked architecture: An urban scene segmentation network based on UNet and ShuffleNetv2’

2.Requirements

• python 3.8
• pytorch 1.11.0
• Cuda 11.3

3.Visualization results of MSA-Net on Cityscapes val set.

*From left to right are input images, ground truth, segmentation outputs.

4.Ablation study results of MSA-Net on Cityscapes test dataset and enhanced PASCAL VOC 2012 val dataset.

Index	Baseline	DLED	ESCC	MSIC	Cityscapes	VOC 2012 Augment	Params
1	✓				63.3	45.5	31.0M
2	✓	✓			65.5	58.2	7.0M
3	✓	✓	✓		72.6	64.2	7.1M
4	✓	✓	✓	✓	73.6	65.3	7.6M

5.Comparisons between different channel depths on Cityscapes val dataset.

*Performed on a single RTX 4090 GPU
*Note that the format {., ., ., .,} represents the channel depth in encoder of MSA-Net, and the channel depth in the decoder and encoder are symmetric. r represents the channel compression ratio.

Model	Channel depth	mIoU	Params	GFLOPs	FPS
MSA-Net	{64, 128, 256, 512, 1024 | r = 1}	74.7	7.6M	43.7	31.0
MSA-Net-Middle	{32, 64, 128, 256, 512 | r = 0.5}	72.0	1.9M	11.4	33.7
MSA-Net-Slim	{16, 32, 64, 128, 256 | r = 0.25}	63.8	0.5M	3.1	36.3

*Results on Cityscapes test set(Anonymous Link)

• UNet(Baseline)
• Baseline + DLED(Stack-Net-More)
• Baseline + DLED_SE(Stack-Net-One)
• Baseline + DLED + ESCC
• Baseline + DLED + ESCC + MSIC
• Baseline + ESCC

*A demo video of segmentation:

• YouTube