08050f40fff41616ccfc3080e60a301a-Supplemental-Conference.pdf

In the main paper, we show the qualitative results on Cityscapes dataset. Here, we display qualitativeresults on ADE20K dataset in Figure 1. The figure clearly shows that our method is better at understandingthedetails. We adopt Grad-CAM [8] to conduct visualization. As shown in Figure 1, we can clearly find our MSCAN showsbetter visualization results. FLOPs are calculated using the input size of 512 512.