Fine-tune result not smooth about fcrn-depthprediction HOT 12 CLOSED

I have solved this issue

from fcrn-depthprediction.

@JackHenry1992, could you please share what the issue and and solution was in the end?

from fcrn-depthprediction.

Hi, @NikolausDemmel , the issue is that I can't get smooth depth for plane (such as low texture places), the result seems strange and coarse. After finetune more layers (such as res5x and layer*) and decrease learning-rate to 0.0001, it has been solved.

from fcrn-depthprediction.

Thanks for the update.

from fcrn-depthprediction.

We think there are a couple of reason why it could look like this:

• you are not ignoring invalid depth pixels during training
• you do not have enough data / augmentations to learn smooth maps
• your loss seems to be very high. Did you forget to divide by the number of (valid!) pixels?

Cheers,
Christian

from fcrn-depthprediction.

Hi @chrirupp , thanks for your advises.
@chrirupp, as you said, my own dataset maybe too small .

Do you have some advises for increasing generalization ability (such as different camera)

• Will it increase the generalization ability of fcrn-network if I use more dataset to train?
• Will it be worse if I use different camera to get images?

from fcrn-depthprediction.

Hi @JackHenry1992,

yes, in general more data is better for generalization.
About different cameras: we found that you will need to correct the predicted depth map by multiplying it with the ratio of focal lengths between new and old camera to get better global scale.
See our recent CNN-SLAM paper https://arxiv.org/abs/1704.03489 for details on this.

from fcrn-depthprediction.

hi, I also train the nyu checkpoint model, but I am stuck with making my own dataset. I don't know how to make label. Do you have any ideas. Thanks~

from fcrn-depthprediction.

Would you share the code for making datasets? Please, thank you!

from fcrn-depthprediction.

The easiest way to create a dataset is to run around with a Kinect or similar sensor. You can try with simulation but that might be quite some effort to render a synthetic dataset.
We did not make the datasets that we use in the paper. They are publicly available (http://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.html)

from fcrn-depthprediction.

Thank you~

from fcrn-depthprediction.

@JackHenry1992

Hi. I am also doing such job, but it did not convergence.
I am so sorry but Can you put your code on GitHub or something else? I want to train it again as reference!
Thank you very much!

from fcrn-depthprediction.