image semantic segmentation package

A image semantic segmentation toolbox (single GPU) contains some common semantic segmentation algorithms. The code is implemented by Pytorch.

Requires

1. pytorch >= 1.0.0
2. python 3.x

Usage

 python train.py --option1 value1 --option2 value2 ...
 For the list of options, please see train.py

Performance

Algorithms	backbone	norm	dataset	batch_size	image_size	Epoch	pixAcc	mIoU
PSPNet [1]	resnet50	bn	ade20k	16	473	120	80.04	41.68
PSPNet	resnet50	bn	ade20k	12	384	30	77.1	38.6
PSPNet	resnet50	bn	ade20k+bk	12	384	30	72.19	35.3
EncNet [2]	resnet50	bn	ade20k	16	480	120	79.73	41.11
EncNet	resnet50	bn	ade20k	8	400	50	77.7	40.3
DeeplabV3 [3]	xception	bn	ade20k	8	384	50	77.6	39.5
DeeplabV3+ [4]	xception	bn	ade20k	8	384	50	77.9	39.8
FCN32s [5]	vgg19bn	bn	ade20k	12	384	50	73.0	31.2

The items with hyperlinks are the experimental results from the original paper

Discussion and details:

In the original paper, authors run their experiments on the standard ADE20k(150 classes, without background). But I regard the background (i.e. labeled 0 in the original mask) as a category and the output dimensionality of the PSPNet is 151 in my code. Therefore, the performance gap mainly comes from three aspects：

1. I add the background class to the dataset, which may lead to category imbalance problems and increases the complexity of the model.
2. Due to limited video memory on a single GPU, I set the batch_size to 12/8 and image_size to 384/400 instead of the parameter settings in the original paper.
3. In addition, the experiments in the original paper used multiple GPUs, which means a larger batch_size can be set to make Synchronization Batch Normalization layers more effective.

TODO

• PSPNet
• ENCNet
• ENCNet+JPU
• Deeplabv3
• Deeplabv3+
• RefineNet
• FPN
• LinkNet
• SegNet
• FCN
• Unet
• Unet++
• DenseASPP
• ICNet
• BiSeNet
• PSANet
• DANet
• OCNet
• CCNet
• ENet
• DUNet

References

[1] Zhao, Hengshuang, et al. "Pyramid scene parsing network." Proceedings of the IEEE conference on computer vision and pattern recognition. 2017.

[2] Zhang, Hang, et al. "Context encoding for semantic segmentation." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.

[3] Chen, Liang Chieh , et al. "Rethinking Atrous Convolution for Semantic Image Segmentation." (2017).

[4] Chen, Liang-Chieh, et al. "Encoder-decoder with atrous separable convolution for semantic image segmentation." Proceedings of the European Conference on Computer Vision (ECCV). 2018.

[5] Shelhamer, Evan, J. Long, and T. Darrell. "Fully Convolutional Networks for Semantic Segmentation." IEEE Transactions on Pattern Analysis & Machine Intelligence 39.4(2017):640-651.