Problem when train from scratch

[Tiam271]

Hello great work @AlessioTonioni and team!
I am trying to use the work for training on custom data which has spare disparity maps (16 bit) as shown below:

I did not modify any code. However, as the training progresses, the disparity predicted by the network has always been like this: (Except for the first frame)

Do you know why? Can you help me?

[AlessioTonioni]

What is the range of the disparities? Are you starting from random initialization or from a set of pre trained weights?

[Tiam271]

The range of the disparities is between 0 and 192.
It is starting from random initialization.
Do you mean that if I start training from random initialization, I should use dense ground-truth?

[AlessioTonioni]

Generally speaking yes it is more stable with full disparities, or at least not very sparse ones.
Have you tried starting from the Flying Things 3D weights?

some additional questions:

• In the 16 bit disparity maps the pixel intensities are mapped to the disparity values multiplied by 256 right? (i.e. as done in KITTI)
• Which value are you using in the GT for pixels without a disparity?

[Tiam271]

• The 16-bit disparity map is exactly the same as in Kitti.
• The value in the GT for pixels without a disparity is setting as zero.

When I start from the pre-trained weights(Flying Things 3D), this problem does not occur.
I understand, this is because the supervision information is too sparse.
Thank you for your answer!

[AlessioTonioni]

I see, so the encoding of the GT seems ok.
Probably as you said you are experiencing collapsing because the supervision is too sparse. I would bet that by playing with the hyperparameters you might be able to train the network directly on the sparse data, but in general I think it's beneficial to start from the F3D weights.
Within the same codebase you can also experiment with Dispnet, even if it is slower it's a way more stable model and you might be able to sucesfully train it on sparse data from scratch (not sure tough)