Dynamic Warping Network for Semantic Video Segmentation

Table 2

Ablation study of the flow-guided convolution on the Cityscapes validation set.

Method	mIoU %
Baseline	73.75
Baseline + warp	74.30
Baseline + FG-Conv	75.02
Baseline + FG-Conv	75.34
Baseline + FG-Conv	75.16

denotes the kernel size of the flow-guided convolution. The bold values denote our method can achieve the best accuracy when s is set to 3.