AdditionalresultsonKITTI2012and2015. 15

Eachconvolutional layer has batch normalization and ReLu activations, except for the output layer. Time(s) GC-Net 2.9M 0.25 GC-Net+CRF 3.4M 0.29 PSM-Net 5.2M 0.31 PSM-Net+CRF 5.7M 0.35 Table 4: Number of model parameters and prediction time on the Sceneflow test set for the four differentarchitectures. The results of this evaluation can be found in Table 5of the supplementary material. The performance gain achieved by adding the CRF is significant for both GC-Net and PSM-Net, which provides strong quantitative evidence in support of the qualitative observations in Figure 5.