search

ialhashim / DenseDepth

High Quality Monocular Depth Estimation via Transfer Learning
https://arxiv.org/abs/1812.11941
GNU General Public License v3.0
1.57k stars 357 forks source link

Question about the inconsistency between train set and test set. #184

Closed Boltzmachine closed 3 years ago

Boltzmachine commented 3 years ago

In the train set, the pixels of depth images are small (approximately 0~1), while in the test set the values are large (> 1000). Can you tell me how to convert them to the same format? Which one is the original depth number? I found many people had the same question but it was has not been solved yet.

ialhashim commented 3 years ago

It doesn't matter, in the end it is normalized from 0 to 1. And then to get the "real" depth values you simply multiply by the maximum depth (in meters, e.g. NYU is 10meters) to get the distance.