Dry Waste Classification Dataset (Metal, Paper, Plastic)

Dry Waste Classification Dataset (Metal, Paper, Plastic) This dataset contains single-item images of dry waste, captured in 2020 using a purpose-built automated sorting device. Each image shows one waste item, photographed from directly above against a consistent background. The dataset covers three material categories: metal, paper, and plastic, totaling 2,043 images (1,636 train / 407 validation). Per-class counts are as follows: metal 755 images (604 train / 151 val), paper 579 images (464 train / 115 val), and plastic 709 images (568 train / 141 val). A small number of glass images (22 total) were also collected during the original project but have been excluded from this release due to an insufficient sample count for meaningful training or evaluation. Baseline validation experiment. To confirm the dataset carries learnable signal, a MobileNetV2 transfer-learning classifier (partially fine-tuned) was trained as a sanity check, achieving 82.3% validation accuracy, well above the 33% chance level for a three-class problem. This result is intended to demonstrate that the dataset trains a real classifier, not as a tuned or state-of-the-art benchmark. Per-class precision, recall, and F1-scores, along with full methodology and discussion, are provided in the accompanying documentation on GitHub. Known limitations. All images originate from a single capture device with a fixed camera position and background, so models trained on this data alone may not generalize well to other cameras, lighting conditions, or backgrounds. Some images depict the same physical object from multiple angles; researchers building their own train/validation splits should consider grouping by object to avoid data leakage. The dataset is moderately imbalanced, with paper underrepresented relative to metal and plastic. At just over 2,000 images, the dataset is modest in size and is best suited as a lightweight benchmark or transfer-learning starting point rather than a large-scale training set. Code and documentation. A baseline training script (PyTorch, MobileNetV2) and full dataset documentation (PDF) are available at: https://github.com/Kamaljust/dry-waste-classification-dataset Citation. If you use this dataset, please cite: Jannati, K. (2026). Dry Waste Classification Dataset (Metal, Paper, Plastic) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.21849964