I use the dataset originally from “Happywhale - Whale and Dolphin Identification” Kaggle competition. For a more structured and easy-to-implement data, I found the resized whale-dolphin images dataset (contributed by “RDIZZL3”). The contributor resized the images to 128×128. It contains 3k training images and 1k testing images with two species: bottlenose dolphin and killer whale.
Want to use the ResNet50 as the backbone of ProtoNet and see its performace.
@article{He2015,
author = {Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun},
title = {Deep Residual Learning for Image Recognition},
journal = {arXiv preprint arXiv:1512.03385},
year = {2015}
}