hog + svm a lot of false negative

asked 2018-09-26 06:36:19 -0600

___SJ___ gravatar image

updated 2018-09-26 07:28:02 -0600

HI! I download dataset from https://gdo152.llnl.gov/cowc/ (15 cm per pixel) and right now have: 6 700 positive tiles (with vehicle) and 7 000 negative (other object&nature) each 120x120 pixel. Then I training HOG + linear SVM, test on learning data with this result:
True Positives: 6409
True Negatives: 6393
False Positives: 607
False Negatives: 309
and when I test my descriptor on Selwyn aerial, I have a lot of false positive and negative alarm (red circle) C:\fakepath\8.png
C:\fakepath\9.png
how to make result better:
-it is possible to make samples less?
-apply filters?
-Change the ratio of positive and negative images in the training sample?

I will be happy with the advice.
Thank you!
C:\fakepath\15.png

edit retag flag offensive close merge delete

Comments

link to dataset looks broken above

also: hog descriptors are not pose / rotation invariant.

(bad choice for your task. humans are usually "upright", that's why it works there)

berak gravatar imageberak ( 2018-09-26 07:08:21 -0600 )edit

what algorithm would you advise? is there any point in writing a rotation invariant modification (seen in the network description)?

___SJ___ gravatar image___SJ___ ( 2018-09-26 07:32:17 -0600 )edit

"seen in the network description" ?? hmm, where ?

berak gravatar imageberak ( 2018-09-26 07:47:01 -0600 )edit
1

www.ijfis.org/journal/download_pdf.ph...
https://lmb.informatik.uni-freiburg.d...

___SJ___ gravatar image___SJ___ ( 2018-09-27 00:38:24 -0600 )edit

did you see the paper linked in the dataset page ? it mentions several cnn architectures to try.

oh, i didn't know FourierHOG before ! maybe that's another idea. (but good luck, porting matlab code)

berak gravatar imageberak ( 2018-09-27 00:50:42 -0600 )edit
add a comment