This repository houses the files and codes used by Henrique Simões and João Meidanis in the research project entitled "Humpback Whale Identification by Its Tails".

Humpback whales (Megaptera novaeangliae) were predominantly explored by the fishing industry between the 1860s and the late 1900s. After the humpback whale fishing was banned, and the species declared endangered, the population started to recover. However, the data of this recovery were still limited. In order to track the population dynamics, researchers use the whale’s fluke shape and markings to identify the individuals. But this process had to be done manually. To address this issue, the Kaggle platform hosted a competition from Nov. 2018 to Feb. 2019 where an automatic model should identify the whales by their tails. In our research project, we aimed at analyzing and reproducing the competition top solutions’ results, as well as at testing improvements that could be incorporated to the best solution.

The files are distributed across several folders which are described next.

• data-analysis: Processed information about the data used by competitors to build their algorithms.
• solutions: Solutions developed by the analysed candidates;
• util: Executables created by us to handle some tasks;
• test: Corrected data (and intermediate files);

This repository contains several branches, which correspond to different changes in the source code.

In this branch (main), the solutions folder contains the top-3 solution code provided by the authors. Small fixes had to be done in order to run the code, such as changing the resources' paths or fixing imports. In order words, this branch contains the code we used to reproduce the creation of each solution's trained model.

The other branches have the following changes:

• {1st,2nd,3rd}-solution: Updated solutions using our datasets;
• gcn: Updated 2nd-solution branch code using Global Contrast Normalization (GCN);
• lcn: Same as above but using Local Contrast Normalization (LCN);
• no-cn: Same as above but using no Contrast Normalization;
• swa: Same as above but using Stochastic Weight Averaging (SWA);

This project was supported by São Paulo Research Foundation - FAPESP under grant #2019/11386-3. We also thank the Institute of Computing/Unicamp for the technical support given during the project execution, which includes the configuration and maintenance of the machine used in our experiments, and acquired with FAPESP's support under grant #2018/00031-7. Moreover, we also thank SAE/Unicamp for the support given.

Opinions, hypothesis and conclusions, or recommendations made in this material are responsibility of the authors, and not necessarily reflect FAPESP's point of view.

The source codes except those in the solutions folder are licenced under the BSD 3-Clause License, and the data resources of this project except those in the solutions folder are licensed under the CC0 1.0 Universal (CC0 1.0) Public Domain Dedication.