Open vaklyuenkov opened 7 years ago
anchor accuracy for rpn proposal network (object/non-object) is likely to get high fast to above 90 percent, but what should take longer is cls accuracy (or class prediction accuracy) -- at least for PASCAL VOC. I got cls accuracy to 90 + only after around 30k iterations
On 30 Nov 2016 14:59, "Klyuenkov Vladimir" notifications@github.com wrote:
Surprisingly high accuracy - more than 90% after only 100 iterations. The same picture on another datasets. It seems, that incorrectly calculated the loss and accuracy. Where may be mistake?
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/alfonsolink/tf_rfcn/issues/4, or mute the thread https://github.com/notifications/unsubscribe-auth/AQ-ORu40nPpPdkJksGv94Amb7w7zFcoAks5rDR7ngaJpZM4K_3i- .
Sorry for bothering you, but i verify this out multiple times. How can it be, that in empty model (without any .ckpt loading) we have rcnn and rpn 90%+ accuracy on test set from first iterations? If I look on results of model at demo, I can see that after thousands iterations accuracy of rpn isn't bed on my pictures, after 50 iterations - roi with best scores not even near goal gt, but as u can see on picture - accuracy , calculated in code, is not bad. Where can be error in calculating accuracy?
And the same can say about losses. bbox_losses starts from small values like 0.08, but on pictures we can see good bboxes only after thousand iterations.
Surprisingly high accuracy - more than 90% after only 100 iterations. The same picture on another datasets. It seems, that incorrectly calculated the loss and accuracy. Where may be mistake?