Can I ask some questions about TOOD/mmdet/core/bbox/assigners/task_aligned_assigner.py ?

[boy-cyh]

line 65: decode_bboxes (Tensor): predicted bounding boxes, shape(n, 80)
Why the shape of bbox here is (n,80)?
I started reading from tood_head.py and I wonder why the shape here is not the (n,4)?

[Chengjian Feng]

You are right. It is (n,4). I made a typing error. Thank you for your reminder.

[boy-cyh]

Thank you for your answer！
And this is associated with a line of code.
the 83 line in the task_aligned_assigner.py : alignment_metrics = bbox_scores ** alpha * overlaps ** beta
In my inference，the shape of “bbox_scores” is (n,1,80) and the shape of "overlaps" is (n ,num_gts).
But the "num_gts" is not equal to 80，It seems that the two cannot be multiplied by "*"。
It's a little confusing to me.

[boy-cyh]

I think if the shape of overlaps is (n, num_gts, 1)，they can be multiplied by "*" to get the alignment_metrics, and the shape of alignment_metrics is (n, num_gts, 80).
I don't know if my understanding is correct about the alignment_metrics.

[boy-cyh]

Sorry, I think of gt_label as None.I understand it.