1
$\begingroup$

ConvNET can easily predict class of an object in an image. My question is, can ConvNET distinguish Pisa Tower from other buildings or Hagia Sophia from other mosquoes easily? If it can, how many training images can be sufficient? Do I need thousands of training images of that specific thing to distinguish it?

(This is a term project recommendation about deep neural networks, so I need to understand its feasibility.)

$\endgroup$
  • $\begingroup$ Is ConvNET some specific architecture, or are you talking about Convolutional Neural Net? $\endgroup$ – DuttaA Sep 12 at 17:27
  • $\begingroup$ @DuttaA No I didn't mean a specific architecture, any approach is welcome. It seems YOLO algorithm suitable, but I'm not sure is this the best approach. Even it's the best, not sure about it's accuracy rate on this subject. $\endgroup$ – Atreidex Sep 12 at 17:39
0
$\begingroup$

To help you understand the feasibility of your project these posts could be a good start: https://datascience.stackexchange.com/questions/13181/how-many-images-per-class-are-sufficient-for-training-a-cnnenter link description here

https://stats.stackexchange.com/questions/226672/how-few-training-examples-is-too-few-when-training-a-neural-network

That being said, the short answer would be it depends. It depends how precise you want to be, what is the difficulty of the task, the infrastructure you have for the training etc.

For the images, you should not worry, you could start with the ImageNet dataset: For construction and buildings: http://www.image-net.org/explore?wnid=n04341686

For mosquees: http://www.image-net.org/synset?wnid=n03788195

You can then use data augmentation techniques to enhance the size of you training set. Here is a library I have used in the past which helped me greatly to achieve this task: https://github.com/aleju/imgaug

Hope that helps!

$\endgroup$

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.