questions about train datasets about learning2adaptforstereo HOT 4 CLOSED

For synthia we rendered our own sequences using the engine. The baseline and focal length of the virtual cameras were selected to closely match the one in kitti.

For Kitti you can either use the raw lidar maps or the cleaned depth maps available here: http://www.cvlibs.net/datasets/kitti/eval_depth.php?benchmark=depth_completion

The code should already take care of discarding not valid pixels (disparity = 0 or disparity > max_disparity) and computing the loss only on the valid ones.

from learning2adaptforstereo.

Thanks for reply~
for the question 1, I mean that you mentioned you use this synthia datasets: http://synthia-dataset.net/ for training offline. So I wanna know the details about converting the depth map to disparity map, because I get the data seems wrong...
Also I wanna ask, the pretrain weights you provided here is just a dispNet weight? or it's a offline trained weight with WAD mode?
Best wishes!
zilch

from learning2adaptforstereo.

As for 1:
we have used the SYNTHIA VIDEO SEQUENCES (CVPR16) available here https://synthia-dataset.net/downloads/ and converted to disparity using the focal length and baseline of the simulated stero system.
depth = (baseline * focal)/disparity

The pretrained weights are for Dispnet traind on FlyingThings3D

from learning2adaptforstereo.

Thanks a lot~
best wishes

from learning2adaptforstereo.