search

yzcjtr / GeoNet

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

About the training strategy #50

Closed zmlshiwo closed 5 years ago

zmlshiwo commented 5 years ago

Hi, I would like to ask about the training strategy of geonet.

  1. To reproduce the result of depth, I use the kitti_raw_eigen split to train the depth and pose networks, 30 epoch.
  2. To reproduce the result of the residual flow, I first use the kitti_raw_stereo split to train the depth and pose networks, 30 epoch, and then I use the kitti_raw_stereo split to train the residual flow, about 200 epoch.

Is the whole process like this? or

  1. Reproduce the result of depth, use kitti_raw_eigen split to train depth and pose networks, 30 epoch
  2. To reproduce the result of the residual flow, I first use the kitti_raw_eigen split to train the depth and pose networks, 30 epoch, and then use kitti_raw_stereo split to train the residual flow, about 200 epoch.

I used depth's training strategy to reproduce the results of depth. However, training the residual flow is a bit worse than the paper results. Can you help me to see which one is wrong?

yzcjtr commented 5 years ago

You should always use kitti_raw_stereo throughout your stagewise experiment for optical flow.

zmlshiwo commented 5 years ago

Ok, that is, to reproduce the result of the residual flow, I first use the kitti_raw_stereo split to train the depth and pose networks, 30 epoch, and then also use kitti_raw_stereo split to train the residual flow, about 200 epoch. All right? Using kitti_raw_stereo split on both the first stage and second stage.

yzcjtr commented 5 years ago

Yes, you should keep the split consistent throughout different stages.