Hi,
I am using the torchmetrics mAP calculator for object detection.
Documentation: Mean-Average-Precision (mAP) — PyTorch-Metrics 1.6.1 documentation
My question is the following:
Lets say I have 20 classes. I know these are required to be 0-indexed. I need a class for background (for images were no objects detected). Should my background class be included? So my background class would be index 0, last class would be index 20.
When model doesn’t detect any classes in a given image, should the predictions dictionary contain a background prediction (label 0, score 0, bbox [0, 0, 0, 0])? Or should it just be empty?
I’ve noticed that if I add a background class and enable per class metrics, I get mAP results for the background class too of course. Obviously the mAP for that class is -1 since it is all wrong detections, but is this correct?
I have read the documentation but cant seem to find this. Maybe its a common knowledge thing so it is just taken for granted.
Thanks.