Animal Action Video Dataset with Heatmap and Silhouette Representations
Description
This dataset presents a comprehensive collection of animal action videos covering 25 distinct animal classes. 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. This dataset can be used for animal action recognition, video classification, explainable computer vision, and related machine learning research.
Files
Steps to reproduce
Raw animal videos were organized by class and preprocessed to standardize format and resolution. Silhouette representations were generated from the videos using background subtraction techniques to capture motion and shape information. Class activation heatmaps were produced using trained convolutional neural network models. The generated data were then organized into training, validation, and test splits for reproducible experimentation.
Institutions
- Pabna University of Science and Technology