Scientific Impact
Evaluating machine learning models for multi-species wildlife detection and identification on remote sensed nadir imagery in South African savanna
Paul Allin, Fadel Seydou, Frans Radloff, Andrew Davies, Alison Leslie
Dataset for: Evaluating machine learning models for multi-species wildlife detection and identification on remote sensed nadir imagery in South African savanna
Paul Allin, Fadel Seydou, Frans Radloff, Andrew Davies, Alison Leslie
Ecosystem Contributions
Optimized and relabeled datasets to support the wider wildlife monitoring community.
The Mission
WildDetect is a comprehensive ecosystem designed for scalable and accurate wildlife monitoring. By automating the transition from raw aerial imagery to detailed census reports, we empower researchers to focus on policy and protection.
A Modular Ecosystem
1. WilData
High-quality, version-controlled data infrastructure. Handles multi-format imports and large-scale tiling.
2. WildTrain
Transform raw observations into specialized AI models. Flexible framework for YOLO detectors.
3. WildDetect
Deploy models in the field. Orchestrates census campaigns and generates sound population counts.
The Workflow
1. Data Preparation
- Raw Annotations (COCO/YOLO)
- WilData Import & Tile
- Processed Dataset
2. Model Training
- Model Architecture
- Training Loop & HPO
- MLflow Model Registry
3. Deployment
- Aerial Image Stream
- Multi-species Detections
- Census Reports