Overview

MammAlps is a camera trap video dataset featuring 8.5 hours of densely annotated animal recordings collected from three sites in the Swiss National Park. Each event was captured from up to three cameras simultaneously. Individual tracks were labeled from the raw videos, and both the species and the behaviors of each track was manually annotated. To better capture the nature of animal behavior, each behavioral label is represented as a pair of a high-level activity (e.g., foraging) and of one or two low-level actions (e.g., grazing). In total, the dataset includes 11 unique activities and 19 unique actions shared among 5 species.
In addition to the dataset, we introduce two benchmarks:
- Benchmark I – Multimodal species and behavior recognition: Using over 6,000 short clips, audio recordings, and background segmentation maps, the objective is to develop models for species, activity, and action recognition.
- Benchmark II – Multi-view long-term event understanding: The objective is to summarize 397 long-term events—seen from up to three views—to identify the recorded species, their activities, meteorological conditions, and the number of individuals.
The MammAlps dataset is publicly released to encourage research in wildlife action and species recognition, action segmentation and localization, long-term video understanding, and multi-animal tracking.
Dataset
For best viewing, zoom in. Video quality has been reduced for faster loading and visualization.
Getting Started
Follow the steps below to get started:
- Download the Dataset: Visit the Zenodo page to download the dataset.
- Documentation (Coming soon): Check out the GitHub repository for detailed instructions and examples.
- Training models (Coming soon): Clone the training code repository and follow the setup instructions to start fine-tuning models.
- Evaluation (Coming soon): Check out the evaluation instructions in the GitHub repository.
If you encounter any issues, feel free to open an issue on our GitHub Issues page.
Upcoming
We are currently processing a new round of data collected in 2024, which will be added in a future version of the dataset. This new batch will increase the number of recordings for species that were moderately seen in 2023, such as hares and foxes.
Additionally, we are actively working on releasing our data processing pipeline to facilitate reproducibility and to spur the development of similar datasets.
Citation and Contact
@article{gabeff2025mammalps, title={MammAlps: A multi-view video behavior monitoring dataset of wild mammals in the Swiss Alps}, author={Gabeff, Valentin and Qi, Haozhe and Flaherty, Brendan and Sumb{\"u}l, Gencer and Mathis, Alexander and Tuia, Devis}, journal={arXiv preprint arXiv:2503.18223}, year={2025} }For any personal inquiry, we invite you to reach out to Valentin Gabeff, Devis Tuia or Alexander Mathis. For any general code or data related questions, we invite you to open a github issue on our GitHub Issues page.