Beyond Humanity: Leveraging Pre-trained Human Video Classification Models for Data-Efficient Multi-speciesWildlife Animal Action Recognition

This paper presents a transfer learning approach for data-efficient video-based multi-species wildlife animal action recognition, using pre-trained models on human action datasets. It bridges the gap between the well-studied human-focused video classification and under-investigated animal action recognition, largely limited by insufficient structured, annotated data across animal species. By leveraging the SlowFast framework, a state-of-the-art architecture for video classification, and conducting on a small sample of the Animal Kingdom dataset, a benchmark on animal action recognition, the paper reveals a notable improvement in the mean Average Precision (mAP) score, with much fewer training data, when fine-tuned on a model pre-trained with Kinetics-400 as compared to training from scratch or utilizing image-based model pre-trained on ImageNet. This research demonstrated the promising nature of cross-domain transfer learning for video classification and has substantial inspiration for advancing the understanding of animal behavior and biodiversity conservation.