Explain parent_object_index

[dishank-b]

For each input box, the (N) tensor parent_object_index contains the index (1-based) in the keep list of the highest-scoring box overlapping this box. This can be useful to group input boxes that are related to the same object. The index 0 represents a background box which has been ignored due to top_k.

Can you please explain the above line from the readme, with an example?

[gdlg]

I’ll try:

Given a list of bounding boxes [a,b,c,d,e,f] with a,b,e overlapping each other and c,f are also overlapping each other and d overlapping none. Assuming that the output of NMS (the highest scoring boxes) is [1,3,5], that is b,d,f are the highest-scoring boxes non overlapping boxes. Then parent_object_index would be [1,1,3,2,1,3] which means that a,b,c,d,e,f are overlapping respectively b,b,f,d,b,f (note that parent_object_index is a 1-based index.

In the case where a box is overlapping multiple high-scoring box, its parent would be the box with the highest score.

I have used with object detection algorithms. If you use NMS to filter duplicates from a list of detection candidates to get a final list of detected objects then you can use parent_object_index to know which candidate belongs to which object.

[dishank-b]

@gdlg Cool, Thanks

[dishank-b]

Also, in the code logically it should be top_k == num_to_keep. So is this the case?

[gdlg]

Not always. You have the guarantee that top_k >= num_to_keep. num_to_keep can be smaller if either the size of your input array is smaller than top_k or your boxes are heavily overlapping each other to the point that you have less non-overlapping high-scoring boxes than top_k.