search

lmb-freiburg / flownet2

FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
https://lmb.informatik.uni-freiburg.de/Publications/2017/IMKDB17/
Other
1k stars 318 forks source link

Problem regarding dataset #115

Closed wzrf closed 6 years ago

wzrf commented 6 years ago

Hi I am training Dispnet model on the kitti dataset, I don't know where I did this wrong. I load the ground truth disparity image, and do the conversion in python: pfm=np.float32(image/np.float32(256.0)) and use the pfm file produced(I use the readPFM and writePFM function in python), and make the lmdb database.

After I train the model, I test it, and I found the float numbers in result are far different from the ground truth, all range in (0,1), although if I use the command line(cvkit tools provided by kitti): sv result.pfm: to visualize the pfm as images, it looks kind of like the ground truth. So where did I do this wrong?

nikolausmayer commented 6 years ago

Does the groundtruth image after pfm=np.float32(image/np.float32(256.0)) look right? Did you read the KITTI groundtruth using KITTI code? If the test results "look" right, they might just be wrongly scaled. If so, I'd suspect the scaling error to happen before training :thinking:

EDIT: Also please check your repository is at least up to date with https://github.com/lmb-freiburg/flownet2/commit/07ced13bbb4da627c269b7cc7a2949de9961b60f

nikolausmayer commented 6 years ago

(closed due to inactivity)