Animal Action Video Dataset with Heatmap and Silhouette Representations

The videos were collected and organized to support research in animal action recognition and video-based computer vision tasks. From each raw video, silhouette representations were generated to capture motion and shape information, and class activation heatmaps were produced to enable explainable artificial intelligence (XAI) analysis.

Heatmaps were generated using multiple convolutional neural network architectures, including EfficientNetB0, ResNet50, and MobileNetV2, to highlight discriminative regions contributing to model predictions. The dataset is systematically organized into raw video data, generated heatmaps, silhouette representations, and predefined training, validation, and test splits. This structure facilitates reproducible experimentation and fair evaluation.