search

mrharicot / monodepth

Unsupervised single image depth prediction with CNNs
Other
2.21k stars 629 forks source link

about the camera pose #197

Open Usernamezhx opened 5 years ago

Usernamezhx commented 5 years ago

camera

I want try to train other data. as you see. I get sample by this way. two camera were nearly vertical. I try to train the net with several pair images to see if it can work. but the loss will be stable at 1.6 + . if this can show that the camera pose is not suitable for the net ???

Usernamezhx commented 5 years ago

loss

Usernamezhx commented 5 years ago

loss loss1

Usernamezhx commented 5 years ago

1 2 this is my data. the cameras were vertical.

pkrush commented 5 years ago

Did you calibrate and rectify the image pairs?

Sorry, I don't know: "If this can show that the camera pose is not suitable for the net?"

Usernamezhx commented 5 years ago

@pkrush thanks for your reply. yeah. I have calibrate and rectify the image. And I remove the lr_loss from the total loss. and get that: 2_disp

mrharicot commented 5 years ago

Hi, Your cameras are not in a stereo configuration I'm afraid. You have to be able to rectify them in order to be able to train with this code