Luca Morandini, Andrea Diecidue, Thanos Petsanis, Enrico Targhini, Georgios Karatzinis, Giacomo Boracchi, Elias B. Kosmatopoulos, Piero Fraternali, Athanasios Ch. Kapoutsis
The DroneWaste dataset is a public collection of aerial images for developing waste recognition models. Each visible waste instance is annotated with a segmentation mask, a bounding box, and a waste category. The dataset contains 4993 images, 5135 annotations, and 20 waste materials. Each category is mapped to a European Waste Code (EWC) to uniquely identify the waste type. The guidelines followed by the researchers and professional photo-interpreters to annotate the images are available here.
The dataset is available for download on the DroneWaste Zenodo repository.
The three object detectors explored in the paper are: YOLOv8, YOLOv12 and Faster-RCNN. The table summarizes performance on the DroneWaste dataset.
| Model | Parameters | mAP@50 | Documentation |
|---|---|---|---|
| YOLOv8x | 68.2M | 38.2% | Ultralytics docs |
| YOLOv12x (turbo) | 59.3M | 38.5% | YOLOv12 repo |
| Faster-RCNN | 41.4M | 36.5% | MMDetection repo |
Follow the setup instructions to create the virtual environments and install the dependencies.
Model training and evaluation on the DroneWaste dataset is performed using a k-fold cross-validation approach. Therefore, one model is trained for each fold and the overall performance is evaluated by combining the results from all folds.
Follow the training instructions to train a model on the DroneWaste dataset.
Follow the evaluation instructions to evaluate a model on the DroneWaste dataset.
This work was funded by the European Union’s Horizon Europe project PERIVALLON – Protecting the EuRopean terrItory from organised enVironmentAl crime through inteLLigent threat detectiON tools, under grant agreement no. 101073952.