question about the distance information in KITTI dataset

[Jichen66]

Hi,

I have a question about the distance information in KITTI dataset. Can we directly use the value of z location(in camera coordinates) as the ground truth distance between the object and camera?

I am looking for your reply. Thank you very much in advance!

Best Regards

[SamSamhuns]

Yes, I think so, I did the training on the raw z location values (without standardizing them) and the prediction works fine. As far as I know, you don't need to standardize the y values (z loc). The original author standardizes the z-loc values and later transforms them back which I think is unnecessary.

However, standardizing the X values (bbox coordinates) is very beneficial. And I think it is better to use the width and height of the bounding boxes rather than using the four bbox coordinates for the distance regression.

[Jichen66]

yes, I also tried the raw z location values before. Btw, what do you mean by standardizing?

[SamSamhuns]

Set the bounding box width and height to 0 mean and 1 standard deviation.

[ArunMathewML]

is it necessary to give zloc value for accurate predictions

[omer-m]

Hy any can tell me about the dataset... because I can't find it on KITTI - Vision Benchmark Suite website

http://www.cvlibs.net/datasets/kitti/index.php