These are chat archives for beniz/deepdetect

9th
Feb 2016
Anurag Phadke
@anuragphadke
Feb 09 2016 22:00
hello - getting my hands into deepdetect. quick question: i have a set of 48x48 sized images, and i need to see whether these images/pattern exist in a bigger image..
i trained a classifier (using NVIDIA's DIGITs) with 15 images that are positive and 15 random images that are negative.
however, the model accuracy is only 40%, any tips on increasing the accuracy? happy to share the training set.
Emmanuel Benazera
@beniz
Feb 09 2016 22:13
You need many more images than that
Anurag Phadke
@anuragphadke
Feb 09 2016 22:14
100? or more? this is all manual curation of data
@beniz bigger question, once i have the dataset, how do i train the classifier? there's one folder that will contain all the correct "images", the other folder will have incorrect..
for incorrect images, can i grab any random image, albeit same size, and use it for training?
Emmanuel Benazera
@beniz
Feb 09 2016 22:17
Look at the tutorials, many info there for training
Anurag Phadke
@anuragphadke
Feb 09 2016 22:17
got it.
was looking for a quick tip here.. already going through the training tutorials
Emmanuel Benazera
@beniz
Feb 09 2016 22:19
Usually images in the thousands
If not there are ways
Anurag Phadke
@anuragphadke
Feb 09 2016 22:19
@beniz where can i learn more about ways? getting thousands of 48x48 images (hamburger UI) on ios
Emmanuel Benazera
@beniz
Feb 09 2016 22:20
It s not totally painless unfortunately
Anurag Phadke
@anuragphadke
Feb 09 2016 22:20
got it
Emmanuel Benazera
@beniz
Feb 09 2016 22:22
Things like dd can make the training curve less steep, but there s a bit of a hill still if you know nothing about it
Anurag Phadke
@anuragphadke
Feb 09 2016 22:41
got it. lemme educate myself first
Emmanuel Benazera
@beniz
Feb 09 2016 22:53
It s a good question though on what could be the best first pages to learn from. I all think about it
Anurag Phadke
@anuragphadke
Feb 09 2016 22:54
@beniz i have done SVM classifiers etc. before. so not completely new to this space..
but that was mostly text based classifier.