Kaggle Happy Whale Competition Summary

🔗 Competition Link

---

🏆 Result:🥉 127/1588 🥉

---

📸 Dataset Preprocessing:

Approaches:

1. Full Body Cropping: This method captures the entirety of the subject.
2. Back Fin Cropping: This technique emphasizes the back fin with an optimal image size of approximately 768x768 pixels.
3. Back Fin with No Background: An enhanced version of the back fin cropping, offering a clearer view.

Key Insights:

• A combination of both the full body and back fin techniques is essential for precise identification.
• Relying solely on back fin images could overlook other significant body features.
• Strategy: Train two distinct models:
  1. One focusing on Full Body Crops
  2. Another emphasizing Back Fin Crops

🌟 Recommendation: For optimal outcomes, train using both cropping methodologies and employ ensemble techniques during the post-training phase.

---

🎚 Class Balancing in MLP Predictions:

To bolster the prediction accuracy of our MLP model, we devised two strategies. The predictions were formatted in a .csv file, with each row presenting 5 potential match candidates ranked by their confidence levels.

Strategies:

1. Reordering based on Repetition and Rarity:

   • If the primary candidate (first element) in a row recurs as the top choice in over 10 other rows, and the subsequent candidate hasn't been the top choice in any other row, their positions are interchanged.
     • E.g., [ind_1, ind_2, ind_3, …] is modified to [ind_2, ind_1, ind_3, …].

2. Prioritizing New Individual:

   • If the primary candidate in a row is tagged as new_individual and the subsequent candidate hasn't been the top choice in any other row, their positions are swapped.
     • E.g., [new_ind, ind_1, ind_2, …] is altered to [ind_1, new_ind, ind_2, …].

Performance Metrics:

• MLP without class balancing yielded a 0.781 LB score.
• MLP with class balancing achieved a 0.809 LB score.

🌟 Recommendation: The incorporation of class balancing strategies significantly enhanced the accuracy of the MLP model.

---