2
$\begingroup$

For my internship assignment I have to implement a proof of concept for an application that is supposed to scan a picture with a carp on it and identify which carp this is. All of the carps that are going to be scanned are known and they all exist in the database, so no new carps are scanned.

Is this possible? I've been searching a lot about this topic and the only thing I found is customvision.ai, but for using this I need to have at least 15 pictures of the same carp per tag, but the client only has 1 picture per carp.

What are your recommendations or do you think this is not possible?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

I'm not sure I understood your question entirely so please correct me if i'm wrong. You're having all the carp tagged, so if I give you a picture of any of them you know exactly which one is in the picture right? If that is the case then you're dealing with a classic classification problem. One simple way of solving such problem will be to use CNN on the input image to extract features, and at the end of the network have N neurons, where each neuron matches one carp. Just apply a softmax over all the inputs to have a probability distribution, and select the highest value (or if the values is highest then a threshold). Regarding the number of labeled examples per class, having only one might be a problem. I would suggest you to look into few shot learning, which are trying to solve this exact problem (training decent models with limited training data).

Improve this answer
$\endgroup$
4
  • $\begingroup$ Yes, it is supposed to know which carp it is. But i'm having trouble training because i need atleast 15 images of 1 tag. But if I read your answer correctly, I am supposed to look at few shot learning. $\endgroup$
    – Jesper V
    Feb 4, 2020 at 10:04
  • $\begingroup$ Yes, there are a lot of recent papers which takes this problem. Also how did you come up with the 15? Have you tried to train a model with the data you're having? As I understood it you don't need good generalization since you're not gonna predict craps that are different from your training set. $\endgroup$
    – razvanc92
    Feb 4, 2020 at 10:10
  • $\begingroup$ If I want to train a tag on customvision.ai I need atleast 15 images per tag otherwise I can't train the tag $\endgroup$
    – Jesper V
    Feb 4, 2020 at 10:18
  • $\begingroup$ I don't know your programming level, but implementing a model like the one i described shouldn't take too much time. $\endgroup$
    – razvanc92
    Feb 4, 2020 at 11:29

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .