#5 Missing bounding boxes problem about carla-2dbbox HOT 9 OPEN

Hi!

Are you using the semantic lidar-based method? If yes, can you make visualization like what you did here?

Maybe it's because the vehicle's position is out of the LIDAR's upper fov range.

from carla-2dbbox.

Hello,
Thanks for your reply. I am using depth camera information. I would like to use the information from depth than semantic LiDAR.

from carla-2dbbox.

Sorry for my late reply. Right, the current version of depth-based annotation still suffer from false negative problem. There are several reasons that can cause it:

• The occlusion filter failed in detecting the car. If this is the problem, then you need to tune the occlusion filter parameters again, i.e. depth_margin, patch_ratio, and resize_ratio. Yes, it requires more works.
• The center point of the car is out of camera FoV. The algorithm use the center point of the car to determine whether the car is inside the camera FoV or not. The algorithm can miss a car if there is only a small portion of the body that appears in camera. I think it can be solved by using a more robust way to perform FoV filter. But currently I don't have time to upgrade the algorithm. Moreover, I don't think this is the problem in your case since the center of the car is clearly visible in the image.

In case you haven't read the details of the algorithm, you can find it here. Thank you.

from carla-2dbbox.

Hello again,

When I tried to collect data at 60fps with the following orientation of camera:
carla.Transform(carla.Location(x=23,y= 0, z=25), carla.Rotation(roll= 0.0, pitch = -90, yaw= 0.0)).
The FOV is set to 69°. I see there is a problem with the bounding box text information.

The problem is:
In an image there is only one object but the bbox txt information contains of two objects where one is out of the image resolution. Why this problem arises and how can we eliminate this?
This problem arises with 90° pitch angle irrespective of altitude, but when altitude increases the inappropriate information decreases. Do you have any idea on this?

I think due to FOV it gets the information and couldn't deal how can we eliminate these. Could you please help me with this?

Below I am attaching the picture along with the bbox txt file for reference.

002470.txt

TIA

from carla-2dbbox.

Hello @MukhlasAdib
Sorry to disturb you. Excuse me. Could you please help me with the above question?

from carla-2dbbox.

Ah right, sorry. It looks like a bug in the algorithm. So I need to look deeper into the problem. I will inform you if I get something, but I cannot promise you it will be fast.

from carla-2dbbox.

@rak7045 Are you using semantic LIDAR or depth camera?

from carla-2dbbox.

Hello,
I am using depth camera information.

from carla-2dbbox.

Yeah, looks like it has something to do with the angle filter and your specific rotation setting. I need some time to check it. But for temporary solution, you can add a simple pixel coordinate filter by yourself before saving the results. Something like

if any(bbox[:,1] < 0) or any(bbox[:,1] > IMAGE_HEIGHT):
    # remove the bbox
elif any(bbox[:,0] < 0) or any(bbox[:,0] > IMAGE_WIDTH):
    # remove the bbox
else:
    # keep the bbox

from carla-2dbbox.