About 2D bounding boxes

Hi,

When it comes to the 2D bounding boxes per image, I have the following three questions:

  1. Where are the annotated images in your dataset coming from? By saying where are the annotated imges from, I mean that if those from multiple cameras, i.e., including front camera, front left camera, side camera, and so on, or just from the single front camera?

  2. If those images are from multiple cameras, is there a way that I can retrieve them camera by camera?

  3. Regarding annotating 2D bounding boxes, are those images rectified or not?

Best regards,
Chris.

@chris I suppose by “2D bounding boxes per image”, you are referring to nuImages

  1. The images are from across the various cameras

  2. One possible way to retrieve the camera from which a given image comes from would be:

    from nuimages import NuImages
    
    
    nuim = NuImages(
        dataroot='/data/sets/nuimages', 
        version='v1.0-mini', 
        verbose=False, 
        lazy=True
    )
    
    sample_idx = 0
    
    sample = nuim.sample[sample_idx]
    
    sample_data = nuim.get(
        'sample_data', 
        sample['key_camera_token']
    )
    
    calibrated_sensor = nuim.get(
        'calibrated_sensor', 
        sample_data['calibrated_sensor_token']
    )
    
    sensor = nuim.get(
        'sensor', 
        calibrated_sensor['sensor_token']
    )
    
    print(sensor['channel'])
    
  3. Yes, the camera images in nuImages are already un-distorted and rectified.