Positions of the camera

#1

Hello,
I was working on a project to get depth from the monocular cameras. There are some ideas which require me to know the exact height of the camera. I was unable to find the exact height, there are transformations given relative to vehicles rear axle, but I had some difficultly with that too. And the height of the rear axle is not available either. Can someone help with this information and provide some guidance?

Also, I would like to know if there are other approaches tried with this dataset to get monocular depth estimate. (Perhaps something like sfm?)

#2

Hi, apologies for the delay in answering your question. The “rear axle” is really the rear axle projected to the ground plane. So you can get the height simply from the camera extrinsics.
I’m not aware of people using SfM on nuScenes, but you could simply use the depth from lidar, and maybe ignore close ranges as parallax will be an issue there.

#3

Hello,
I am also working on similar task, where I need to calculate the distance of the Traffic light from the vehicle. I tried to apply sfm, but the results obtained were not good. (I am not yet very sure if I applied it correctly as I am still trying on it.) I think the reason for that is, more uncertainity in measurement due to movement in forward direction.
image
I think, Lidar is not a solution for this usecase because Lidar points are not available above certain height.

Did you find any other way to calculate the depth?

Can someone suggest me, if there is any other way to calculate the depth of Traffic Lights from Vehicle using monocular images. Thanks in advance.

#4

Why not simply regress the depth in your network? (essentially doing 3d object detection) That should work “well enough”.

#5

Hello Sir, Thank you for your reply. If I am not wrong, for that I would need ground truths to train my model. If possible could you please elaborate it a bit.
Also would it be possible to know the camera angle of Front Cam at which it is mounted. Thank you Sir.

#6

You can compute the depth from the position ground truth and get the camera angle from the extrinsics.
Please look at the tutorial (https://www.nuscenes.org/nuscenes#tutorials) and the other questions in this forum and on Github.