Questions. about object-detection-metrics

[rafaelpadilla]

Dear @VinayakaPrabhu,

Answering your questions:

1. It is hard do say which threshold is the best when evaluating mAP. It depends on how many classes you have and their particular AP. The number of TP and FP can help, but it is a trade-off between how many wrong detections you allow to have in order to obtain the TP. In a real system you should evaluate how critical your results will be if you increase the amount of FP detections.

2. I don't know Iamr. Do you have a paper or a reference to it? Have you googled it?

3. IOUthreshold is the threshold that will define if your detection is a TP or a FP considering the area of intersection with the object (it is explained in the README). If your IOU threshold is very strict (high values) your bounding boxes must be very close to the ground truth ones. If your threshold is very low, you are accepting a relative small area of intersection between your detection and the object (bad detections). Usually the IOUThreshold is set to 50% to 75%. But it all depends on the purpose of your application.
   The term "threshold" itself is usually applied to the confidence of the detections. Usually object detectors retrieve bounding boxes, classes and the confidence values. The confidence value means how confident the detector is about the class of the object.

I hope it helped.

[VinayakaPrabhu]

Thanks a lot.

is there any way to plot these two curves in the result?

Or anything like this? So. That a best threshold or best mAP etc can be extracted?

[rafaelpadilla]

Dear @VinayakaPrabhu,

Those curves are usually used to evaluate models used for classification. This is not the goal of our repository. Here we use a different metric for object detection only.

Regards,
Rafael