the training of depth predictor

[minghongss]

How is your depth predictor trained? What is the code for training? What training set did you use?

[jahaniam]

To provide depth prediction in the initialization of map
points in CNN-SVO, we adopt the Resnet50 variant of the
encoder-decoder architecture from [7] that has already been
trained on Cityscape dataset. Next, we fine-tune the network on stereo images in KITTI raw data excluding KITTI
Odometry Sequence 00-10 using original settings in [7]
for 50 epochs

We used https://github.com/mrharicot/monodepth code for training although since then the depth neural networks improved a lot and you can use any other arbitrary network.