search

nianticlabs / monodepth2

[ICCV 2019] Monocular depth estimation from a single image
Other
4.18k stars 964 forks source link

Monocular depth estimation #167

Closed sdpscnc closed 4 years ago

sdpscnc commented 4 years ago

Hi, thanks for the amazing work @mrharicot and team! I used KITTI's very small data of about a few hundred frames to verify that the program can start training for my next work. I have not made major changes to the program, except to comment out the depth_gt part. But when I started training, the loss value never dropped, but fluctuated around 0.2. I changed the batch_size and learning rate, but the loss value did not change. Could you please guide me where may there be a problem? Thank you very much!

mrharicot commented 4 years ago

Hi, Did you change anything other than the number of training images? I have observed that monodepth2 can fail to train on very little data

sdpscnc commented 4 years ago

Thanks for your answer! I have not changed anything else, except to use very little data and comment out the depth_gt part.

mrharicot commented 4 years ago

Then it's very likely coming from the lack of training data.

sdpscnc commented 4 years ago

I'm sorry, I'm very stupid! I have made a problem in using KITTI data. My data is not same to 2011_09_26_drive_0001_sync. Now, I can train the program with little data and can begin my next work. Thanks for your amazing work!

sdpscnc commented 4 years ago

In addtion, whether the training data is continuous frame image or not?

mrharicot commented 4 years ago

Hi, Glad to hear you got it to work! KITTI has continuous frame data at 10Hz, it is important that there is enough camera motion between consecutive frames to obtain good depth results.