search

ialhashim / DenseDepth

High Quality Monocular Depth Estimation via Transfer Learning
https://arxiv.org/abs/1812.11941
GNU General Public License v3.0
1.58k stars 354 forks source link

According maxDepthVal inloss.py #58

Closed resindraburiza closed 5 years ago

resindraburiza commented 5 years ago

Hello, good evening. Firstly, I would like to thank you for releasing your code. It could be a lifesaver for me to learn more about learning-based method.

I want to ask about maxDepthVal in loss.py's argument input. In the NYU_BasicAugmentRGBSequence it is stated that maxVal is 1000.0 which I believe is 1000 cm. Is it correct? But, for calculation of SSIM, why do you use 1000.0/10.0 as maxDepthVal? I wish that you would like to help me understand this even though this might be just a very simple question. Thank you very much.

Best, Resin

ialhashim commented 5 years ago

Looking at https://github.com/ialhashim/DenseDepth/blob/master/data.py#L106

You can see that the data is 'normalized' using the maxDepth / depthValue. So the maximum is not 1000 anymore but rather 1000 / minimum expected depth.

resindraburiza commented 5 years ago

thank you for very fast reply. I really appreciate it. now I see that in the training, the inverse depth is used. By the way, from where did you get the minimum expected depth = 10 information? Thank you so much.

ialhashim commented 5 years ago

I think it is mentioned in the NYU original page

resindraburiza commented 5 years ago

Thank you! I'll check it out soon.