mukhlasadib / carla-2dbbox Goto Github PK

Hello,

I had some issues with your filter and wanted to achieve nearly 100% accuracy. Thus, I used a similar approach to yours. I wanted to get images from drones, like the VisDrone dataset. My method involved spawning cameras randomly above the map at heights between 50-80 meters. I used a function to get the bounding box of each vehicle from all camera angles and then performed ray casting. I identified the closest ray from the camera, which returned the location and type of the object. If the type was a vehicle, we were 100% certain it needed to be included.

If you have many types and may miss something, you can use distance: calculate the distance between the closest ray and the camera, compare it with the end of the ray_point (which is the vehicle or your specified type), and if they are almost equal, you should add it. Finally, by removing the vehicles based on angle and distance as you have already implemented, I was able to create a perfect simulated dataset.

If you like the idea, I can send you some code to test it yourself and maybe add it as a new function in your code. I am using carla 0.9.15 and this version worked perfectly. Here is a sample of a generated image with the annotations

Hello Adib,
This is not an issue, I just wanted to ask you about the occlusion filter but I cannot find your email to ask. I am sorry for that.
I would like to ask how bad it is when we only compare the actual distance to one pixel in the image, and how comparing the actual distance with all pixels within the bounding box would help to solve the problem?
Thank you so much and hope to hear from you soon.
Best regards,
Luu Tung Hai

I am trying to extract data from four different cameras attached to the same ego vehicle but only Camera1 gives out bboxes the rest of cameras: cam2,cam3, and cam4 return empty bboxes.

even though there are vehicles in the image. I can see the cameras are spawned properly as they output images with vehicles in it but the bboxes are empty.

I am using auto_annotate function

Hello,

Thanks for sharing your great efforts, may I ask which version of CARLA your module was built?

Hello again,
I am collecting data for a CNN network to learn. As already mentioned in #5 about missing Bboxes in the field of view has been solved. I see the problem still persists.

I attached a sensor to a car. The Altitude (z-axis) and the Picth angle of the camera changes according to the requirements.
While collecting the data the missing bboxes occurs. I am attaching the picture for reference

The sensor which is attached to the object is stationary due to traffic signal. Then, a car has appeared in the frame as shown above.

The above image is the frame where the car entered and after for some frames the object has bounding box to it.

While the object is at the end i.e while leaving it has no bounding box to it. This problem has appeared for each requirement we collected.

Do you have an idea on this how to solve the problem?

TIA

Regards
Ravi

Replace literal 100 with max_dist

Thanks for your amazing work.

I have noticed two main problems in your module:

1. The first one is the shifted boxes as shown below
   
   I think this mainly because of CARLA server itself, what do you think?

2. The second problem is missing Bboxes due to the area that you have chosen becomes out of field of view.
   I think we could easily solve it by using the segmentation mask instead of generating smaller box from the original one.
   

Could you clarify why you wrote in test_semantic_lydar

            # Check whether it's time for sensor to capture data
            if time_sim >= 1
                       ...
time_sim = time_sim + settings.fixed_delta_seconds

Why do you capture data only every 1 second?