0
$\begingroup$

I tested py-faster-rcnn and found that if I don't use pretrained weight then mAP is 0 in testing.

I hope that I should see some value eventhough not very high.

With pretrained model, mAP was above 99%.

Why so difference and how can I solve the issue?

With pretrained weight

AP for vertical = 1.0000
AP for horizontal = 0.9500
AP for plate = 1.0000
Mean AP = 0.9833
~~~~~~~~
Results:
1.000
0.950
1.000
0.983
~~~~~~~~

Without pretrained weight,

AP for vertical = 0.0001
AP for horizontal = 0.0000
AP for plate = 0.0000
Mean AP = 0.0000
~~~~~~~~
Results:
0.000
0.000
0.000
0.000
~~~~~~~~
$\endgroup$
  • $\begingroup$ I changed the convolution layers with xavier weight initialization. I saw some improvements. But not significant. $\endgroup$ – batuman Dec 6 '18 at 1:47
0
$\begingroup$

I have tested with all weight initializations and xavier is the best. We need to initialize weights if not the network doesn't converge.

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.