search

mrharicot / monodepth

Unsupervised single image depth prediction with CNNs
Other
2.21k stars 628 forks source link

Bad results while training on EuRoC dataset #234

Closed Beniko95J closed 4 years ago

Beniko95J commented 5 years ago

Hi, I would like to train your monodepth model on the EuRoC Machine Hall dataset, which is more difficult than the KITTI dataset. I have previously stereo rectify the images for training, but the trained result is not good. In fact, the loss is fluctuating. Here is the last part of training logs:

Screenshot from 2019-07-23 14-38-02

I have tested the trained model on a single image, and as expected it does not work properly:

1 1_disp

Is there any suggestion about training monodepth model on such a difficult dataset? Thanks!

mrharicot commented 5 years ago

Hi, Can you share a few examples of the stereo rectified images?

Beniko95J commented 5 years ago

Thank you for your reply!

Here are some examples of my stereo rectified images, in fact I found they have a very short baseline:

  1. 1403636734763555584_left 1403636734763555584_right

  2. 1403636987151666432_left 1403636987151666432_right

mrharicot commented 5 years ago

Hi, These images are not rectified. monodepth assumes rectified images and will not work otherwise. Things should only move horizontally between rectified images.

Beniko95J commented 5 years ago

You are right. After I checked and modified my codes, I get a relatively good result. Thank you very much!

1 1_disp

By the way, may I ask what is the difference between model-40000 and model-71600 in my log directory? The prediction from these two models are quite similar.

Screenshot from 2019-07-28 11-48-49

Thanks!

zaher88abd commented 4 years ago

Hi @Beniko95J, may I ask you How many images you used to train the model?