CODEWITHSUNDEEP
tensorflowobject-detectiontensorboard
SM54
SM54

Reputation: 41

mAP (Mean Average Precision) in Tensorflow Object Detection API

I read somewhere that the mAP metric shown in tensorflow object detection API is different than the mAP given in the model zoo (where ssd inception v2 has mAP of 27 on MSCoco dataset) . When am training my own dataset, I have no reference, as the plot for mAP is between [0,1] while model zoo has mAP higher of 27.

Questions 1. Is the range of mAP =[0,1] in the tensorflow Object Detection API? 2. What mAP should we aim for while training our own dataset?

Upvotes: 2

Views: 4139

Answers (1)

Dmitry
Dmitry

Reputation: 91

Mean Average Precision (mAP) should always be in range [0,1], because pipeline of computing mAP is the next:

  1. Compute precision for each class p1,p2,..,pn (precision, as we know, is in range [0,1])
  2. Compute mean precision of average precisions from previous step mean(p1,p2,..,pn)

So, it proves, that mAP must be in range [0,1]

Upvotes: 3

Related Questions