Results:

Performance on Cityscapes (19 classes, 1024 x 2048)

Training: 768x768 random crop
validation: 1024x2048

pytorch version

Model	Batch Size	FLOPs	train/val OS	mIoU	overall accuracy	mean accuracy	FreqW accuracy	time (ms)
DeepLabV3Plus-MobileNet	16	135G	16/16	0.721	0.952	0.800	0.913	38.06
DeepLabV3Plus-ResNet50	16	N/A	16/16	0.763	0.957	0.840	0.921	22.02
DeepLabV3Plus-ResNet101	16	N/A	16/16	0.762	0.959	0.838	0.924	53.82

tensorflow version

Model	Batch Size	FLOPs	train/val OS	mIoU	overall accuracy	mean accuracy	FreqW accuracy	time (ms)
DeepLabV3Plus-ResNet18	8		16/16	0.648	0.9421			304.37
DeepLabV3Plus-ResNet18	4		8/8	0.649	0.945			352.53

Errors:

1. different size error:

    Invalid argument: padded_shape[0]=212 is not divisible by block_shape[0]=36
   

   Solution:

    set the same size of image for training and validation
   

2. number of samples % batchsize = 0

    Invalid argument: slice index 17 of dimension 0 out of bounds.
   

Reference

1. Atrous Convolution

   https://towardsdatascience.com/understanding-2d-dilated-convolution-operation-with-examples-in-numpy-and-tensorflow-with-d376b3972b25

2. Deeplabv3+

   https://towardsdatascience.com/review-deeplabv3-atrous-convolution-semantic-segmentation-6d818bfd1d74

   https://sh-tsang.medium.com/review-deeplabv3-atrous-separable-convolution-semantic-segmentation-a625f6e83b90