SortWaste | Industrial Waste Sorting Dataset

Introduction

SortWaste is a densely annotated dataset for waste detection created to support research in computer vision for waste management. Collected at an industrial sorting facility, it captures real-world conditions such as cluttered scenes, overlapping objects, and deformed materials, reflecting the visual complexity encountered in operational recycling lines.

The dataset contains high-quality labeled images across 8 waste categories and a range of environmental conditions. By providing accurate and consistent annotations, SortWaste aims to accelerate the development of automated sorting systems and contribute to more efficient recycling and resource recovery.

Annotated examples from SortWaste

Data Collection Process

Top-down capture setup at the sorting line

Industrial Environment

Collected at a Mechanical-Biological Treatment (MBT) facility, the dataset captures waste as it appears on a real conveyor belt:

Top-down perspective
Real cluttered industrial scenes

Dataset Statistics

Eight types of materials were annotated

PET (Polyethylene Terephthalate): rigid, clear or green items, typically bottles, jars, and beverage containers.
HDPE (High-Density Polyethylene): opaque, colored, denser plastics such as yogurt cups and bottles for food, hygiene products, detergents, or alcohol.
ECAL (Liquid food cartons): multilayer packs with at least 75% cardboard, used for liquids like milk or juice.
PET Oil: PET containers for edible oils; treated as a PET subtype due to typical contamination.
Mixed Soft Plastic: flexible, compressible plastics (wrappers, chip bags, plastic bags).
Mixed Rigid Plastic: rigid plastics not classified as HDPE, often transparent (molded packaging, boxes, hard containers).
Cardboard: corrugated or flat packaging used for storage and transport.
Metal: steel or aluminum packaging such as cans and food tins.

Object counts per class and dataset split

Split	HDPE	ECAL	PET	Mixed Soft Plastic	Mixed Rigid Plastic	Cardboard	Metal	PET Oil	# Images	# All Objects
Train	16803	13649	11976	9077	7066	1524	945	802	3705	61842
Validation	4972	2552	2108	1443	1120	425	277	168	780	13065
Test	3269	3026	2722	1817	1230	207	215	132	776	12618
Total	25044	19227	16806	12337	9416	2156	1437	1102	5261	87252

The ClutterScore Metric

We define ClutterScore to gauge the scene’s hardness level using proxies that affect visual complexity:


            ClutterScore = α·H_c + β·N + γ·H_s + δ·O

H_c (Class entropy): measures class diversity/uncertainty in the scene.
N (Object count): number of detected objects in the frame.
H_s (Size entropy): variability of object sizes (small vs. large objects).
O (Bounding box overlap): degree of occlusion/crowding via overlap between boxes.

For implementation details and parameter settings, please refer to the paper or code.

Complexity Levels Visualization

Very Low Clutter

Low Clutter

Medium Clutter

High Clutter

Access the Data

The dataset is available in multiple formats:

Original videos (.mp4);
The complete dataset before splitting;
The dataset with the respective splits, available in COCO and YOLO formats;
The dataset containing only plastic classes, also available in COCO and YOLO formats;

Citation

@misc{inácio2026sortwastedenselyannotateddataset,
      title={SortWaste: A Densely Annotated Dataset for Object Detection in Industrial Waste Sorting}, 
      author={Sara Inácio and Hugo Proença and João C. Neves},
      year={2026},
      eprint={2601.02299},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.02299}, 
}

SortWaste: A Densely Annotated Dataset for Object Detection in Industrial Waste Sorting