Can't converge when trainning using TensorFlow about fcrn-depthprediction HOT 6 CLOSED

Dear chenynCV,

it appears that you use the labelled dataset only and do no augmentations. We found that sampling the 400GB of the raw dataset plus augmentations was crtitical to the training. Check the paper for details.

Best,
Chris

from fcrn-depthprediction.

Year, I have used the labelled dataset only indeeed. I think this process is like fine-tune. It should converge if I use the model right.

from fcrn-depthprediction.

Ah, I see... You are loading our pre-trained model. Maybe the problem is that you do not divide by the number of valid pixels in your L2 loss?

from fcrn-depthprediction.

Thanks for your reply. I will try it later.

from fcrn-depthprediction.

Hi @chenynCV this code that you're using for training does not implement the reverse huber loss correct?

from fcrn-depthprediction.

@chenynCV Hi, i am trying to implement this paper, but the depth maps are blurred, wondering what i am missing,
any suggestions are appreciated ...

from fcrn-depthprediction.