Model about stereo version about monodepth HOT 9 CLOSED

@mrharicot , I found another unsupervised method in herehttps://github.com/tinghuiz/SfMLearner. and he also use dispnet network, and i solved the stereo model training problem according to this method.

from monodepth.

Hi,

I did not figure out why the stereo model became unstable.
It never happened in our previous Torch implementation.

The only real change between both versions comes from the implementation of the bilinear sampler and the deconvolutions which have been replaced with resize-convolutions.

You could also try adding some regularization to the network and see if it helps.

from monodepth.

Thanks for your reply. I will try to solve the problems of stereo model according to your advices.

Another question:
Mono version thrained with stereo rectified image pairs, the main loss is the image reconstruction loss, but the disparity exist only when the stereo image pairs existed. At same depth, the disparity will change if the baseline or focal length changed.
The model is trained with image pairs have same baseline, focus length. If I use another camera which have different baseline or focus length to achive stereo image pairs. I doubt that this mono model won't work.

from monodepth.

Hi,

I am not quite sure of what you mean by "the disparity exist only when the stereo image pairs existed".

If you test on images which have a different focal length the result will probably be less accurate.
However you can fine-tune a pretrained on data with a different baseline, which is what we did in the paper. You could also possibly mix both datasets, however it might be confusing for the network as the same depth will result in a different disparity. I suspect that the network would overfit and learn to identify which dataset the image is from.

from monodepth.

Hi,

By "the disparity exist only when the stereo image pairs existed" , I means that disparity exist only when scene is imaged more than one viewpoint, and may be different if the position changed betweent different viewpoint.
So , I think the main Limitations of your methods is cannot used to different cameras.

Another question, I did not recurrence results of stereo model in your paper, can you tell me the params setting when you are training the stereo model?

thanks a lot!

from monodepth.

Indeed the current model only supports one camera at a time, nothing is stopping it from learning from different cameras/baseline at the same time but I predict the results won't be great.

The stereo model was trained with the exact same parameters as the mono version.
We found out that the model became unstable after ~12 epochs, which somehow did not happen in our previous Torch implementation. We thus stop training after 12 epochs.

from monodepth.

Hi, @mrharicot , I also want to evaluate this method for indoor scenes, can you provide some pretrained model and datasets for indoor scenes .
@Aryayay , can you share your results in indoor_scenes predicted by monodepth?
It will be very grateful if you can upload your indoor_dataset, cause I can not find suitable stereo indoor dataset.

from monodepth.

I do not have models nor datasets to share for indoor stereo unfortunately.

from monodepth.

@JackHenry1992 , my indoor dataset is quiet simple, just use a stereo camera and capture some video.

from monodepth.