This is a visual recognition model designed for fine-grained classification of marine species. The model leverages multi-scale contextual inputs to capture interactions between the object and its habitat. In addition, it encodes the hierarchical structure of marine taxonomy into the feature space, allowing the model to reflect hierarchical relationships in its predictions. By integrating instance and environmental features with taxonomic structure, the model achieves improved performance in fine-grained marine species classification.
While we employed standard classification accuracy for training, the primary evaluation metric was hierarchical distance, defined by the number of hops between the predicted label and the ground truth in the taxonomy tree. The final model attained a public distance score of 1.62 and a private distance score of 1.45 on the official evaluation leaderboard.
