search

yzcjtr / GeoNet

Code for GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose (CVPR 2018)
MIT License
726 stars 181 forks source link

training problem #57

Closed baorui123 closed 5 years ago

baorui123 commented 5 years ago

when i try to train the depthnet, the loss cannot go down(due to the rigid_loss part), it is between 1.0-0.2,soon higher and soon lower i use the data from my own video, so the video should meet what requirement ？ After data preprocess, the concat picture is how like?? help! (i cant get kitti dataset)

yzcjtr commented 5 years ago

Hi, the camera recording the video should be moving. To make the warping loss meaningful, you should select an appropriate sampling rate of the video, i.e. try to keep the area of common contents between consecutive frames reasonable.