CODEWITHSUNDEEP
tensorflowmachine-learningimage-classificationmobilenet
Martin Cremer
Martin Cremer

Reputation: 5592

training of image classification with mobileNet: should it be exactly 1 specimen per image?

I am training a Mobilenet neural network to identify animal species. I wonder: For a good result is it important that every image contains just exactly 1 individual? Or can i also train it with images containing multiple individuals of the same species?

Upvotes: 0

Views: 203

Answers (1)

Gerry P
Gerry P

Reputation: 8112

you will probably get the highest accuracies is you use images with just one example of the species in the image. However now the model is trained that way and if you submit a test set images that have multiple examples of the same species in the image it might not perform very well. So if you want your model to cover the broadest range of potential images I would train it on images that have both cases, that is one example of the specie per image and several examples of the specie in the image. You will get even a more accurate but less general model if you crop your training image such that the majority of the pixels in the image are of the specie.

Upvotes: 1

Related Questions