OpenDriveLab / LightwheelOcc

LightwheelOcc: A 3D Occupancy Synthetic Dataset in Autonomous Driving
Apache License 2.0
67 stars 3 forks source link

How to use depth GT for training? #1

Closed getterupper closed 3 months ago

getterupper commented 5 months ago

Hi, thanks for your excellent job. I noticed that the depth GT provided in the dataset is not the actual depth of the scene but values between 0-254. Could you please share about how do you use the depth GT for training?

sephyli commented 5 months ago

Sorry for the inconvenience. We forget to include the information to docs.

Depth maps are stored as uint8 PNG images, where a value of 255 signifies an invalid pixel (i.e., out of range). For valid pixels, the depth in meters can be calculated by converting the uint8 value to a float and multiplying by $1000.0 / 255.0$:

$$ depth(u, v)=I(u, v) \times \frac{1000.0}{255.0} $$

We have observed that the transformation process introduced some numerical errors. We plan to upload a version with improved accuracy shortly.

ZM-Zhou commented 4 months ago

Hello, I have downloaded the depths from huggingface and I found that the depth maps have many stripes. it seems like they are out of the range of int8. How can I convert them to true depths? 1710168978 394248

sephyli commented 3 months ago

We have released a new version of depth image with more accurate precision. Please refer to Getting Started to download the data. The usage of the data can be found in toturial.ipynb.