Dynamic Warping Network for Semantic Video Segmentation

Table 3

Ablation study of the prediction consistency loss on the Cityscapes validation set.

Method	mIoU %
Baseline	75.34
Baseline +  + w/o mask	75.13
Baseline +  + w/mask	75.46
Baseline +  + w/mask	75.62
Baseline +  + w/mask	75.44

denotes the amplitude of the difference between high warping error and low warping error. The bold values denote our method can achieve the best accuracy when delta is set to 2.