Issue with output of NN classifier

Azure / ObjectDetectionUsingCntk

Fast R-CNN Object Detection on Azure using CNTK

128 stars 61 forks source link

Issue with output of NN classifier #6

Closed AkshayAgarwal007 closed 7 years ago

AkshayAgarwal007 commented 7 years ago

The mAP in case of NN classifier seems to be very low when I tried it out.

nn_issue

The output on the five test images looks something like this:

nn_0win_20160803_11_28_42_pro nn_1win_20160803_11_42_36_pro nn_2win_20160803_11_46_03_pro nn_3win_20160803_11_48_26_pro nn_4win_20160803_12_37_07_pro

AkshayAgarwal007 commented 7 years ago

I tried the entire process twice and it is yielding the same result.

PatrickBue commented 7 years ago

This is a very small dataset, and hence just training a SVM on the output of the NN is expected to work better. You could try tweaking the parameters (esp. the learning rate) for the NN training to see how much the accuracy improves.