To download the models :
bash download_model.shYou will download 8 models, including the segmentation network that we used to remove the sky in the 3D tasks :
-
ade20k_resnet50dilated_encoder.pth : encoder part of the pre-trained segmentation network ResNet50dilated + PPM_deepsup
-
ade20k_resnet50dilated_decoder.pth : decoder part of the pre-trained segmentation network ResNet50dilated + PPM_deepsup
Our models trained on the MegaDepth training set :
-
MegaDepth_Theta1_Eta001_Grad0_0.807.pth : trained on MegaDepth training set, all the results without fine-tuning in the paper is with this model.
-
MegaDepth_Theta1_Eta001_Grad1_0.774.pth : trained on MegaDepth training set with regulazing the gradient of the predicted flow. We generate visual results with this model.
Our models fine-tuned on the test set :
-
MegaDepth_TestFT.pth : trained on MegaDepth then fine-tuned on the MegaDepth test set for the correspondences tasks.
-
RobotCar_TestFT.pth : trained on MegaDepth then fine-tuned on the RobotCar test set for the correspondences tasks.
-
KITTI_TestFT.pth : trained on MegaDepth then fine-tuned on the KITTI test set for the optical flow tasks.
Finally, we also provide Moco features of ResNet-50, which is originally available here