Three Ways to Improve Semantic Segmentation with Self-Supervised Depth Estimation

Benchmark Model Rank Results
semi-supervised-semantic-segmentation-on-cityscapes-100-samples-labeledSegSDE (MTL decoder with ResNet101, ImageNet pretrained, unlabeled image sequences)#5Validation mIoU: 62.09%
semi-supervised-semantic-segmentation-on-cityscapes-12-5-labeledSegSDE (MTL decoder with ResNet101, ImageNet pretrained, unlabeled image sequences)#22Validation mIoU: 68.01%
semi-supervised-semantic-segmentation-on-cityscapes-25-labeledSegSDE (MTL decoder with ResNet101, ImageNet pretrained, unlabeled image sequences)#20Validation mIoU: 69.38%