1
$\begingroup$

I want to train a ann for object detection and my first task is to label my images.

These classes can be partially covered by other classes. Sometimes the middle part is not visible (left/right or top/bottom part is visible), sometimes only one end is visible. I have tried to simply visualize it in an sample image:Sample image

Blue is my target class I have to label. Red is another class or something random. Whats the right way? In the left case I would label the whole area - including parts of the red object. In the right case I would label from the top until the red area begins.

Thanks!

$\endgroup$

1 Answer 1

1
$\begingroup$

You label the images in the same way that you would like them to be detected. You will need to be consistent - if working with others you should agree how to deal with obscured objects carefully.

If your detected objects are nearly always fully in frame, then you might prefer to always mark the full extent of the object (or a best guess), including the obscured edges of a detected object, because then size and pose cues from the visible parts of the object will be more consistent.

You may also have situations where the centre and one edge of an object is obscured, but prefer the detection to properly identify the centre of the object, not the centre of the visible parts. This is probably the most common use case for detection (as opposed to segmentation) but ultimately you care about how you should interpret the detections, and for that you prepare the ground truth as you need it.

In your examples, I would therefore label the two blue objects identically, ignoring the obstruction as best I can. At least I consider that the default. However, if you want the detection to highlight only visible areas in certain cases, then your suggested approach should be fine if applied consistently.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .