Negative Data for DetectNet?

Hi, I am training object detection model to deploy on Jetson TX2.

I wonder if negative data are necessary to train DetectNet.

E.g., do I have to annotate bounding boxes for ‘mouth’ to detect ‘eyes’?

I understood as classification requires negative data.

How about object detection?


No negative data is needed.
Pass the location of your target is enough.

For detection problem, the positive and negative samples are decided by MAP, the overlapping ratio between prediction and ground truth data.

That is:
MAP > threshold : positive
MAP < threshold : negative