This paper discusses the application of deep learning for individual bird recognition.
The manuscript is written clearly, and there is no major issue with understanding the details. However, several issues must be addressed before considering this paper for publication.

1- The authors said that their data collection procedure could work for wild birds. However, their method requires fitting the birds with pit-tags, and it is not clear for me how this could work for wild birds.

2- From the text, it is not clear how does the data collection procedure work if multiple birds were present at the feeders.

3- Although the authored have gathered lots of images, the size of the dataset is significantly smaller than other deep learning datasets such as ImageNet or Snapshot Serengeti with millions of pictures. VGG19 is one of the more massive deep learning architectures with hundreds of millions of parameters. How the authors made sure that no overfitting happened? Why didn't the authors use smaller architectures such as ResNet-18?

4- The authors have used an image classification model and cross-entropy loss for individual recognition. There are other deep learning methods specially designed for individual identification, such as triplet loss. Why didn't the authors use those methods?