Dynamic Warping Network for Semantic Video Segmentation

Table 1

Ablation study of our DWNet on the Cityscapes validation set.

Warp		FG-Conv	FRM		mIoU %
					73.75
					74.30
					74.87
					75.34
					75.25
					74.76
					75.62

Warp denotes the original warping operation. and denote the feature consistency loss and prediction consistency loss, respectively. FG-Conv denotes the flow-guided convolution. FRM denotes flow refinement module. The bold values denote our method can achieve the best accuracy compared with other methods.