A synthetic outdoor waste image dataset with YOLO-format annotations for object detection
Description
This dataset contains 6,000 synthetic RGB images for outdoor waste object detection, annotated in YOLO format. The images represent realistic outdoor environments such as streets, parks, roadsides, and public open spaces, designed to support research in environmental monitoring and intelligent waste management. The dataset includes 10 waste categories: plastic, paper, cardboard, metal, glass, organic waste, battery waste, e-waste, cloth, and other waste. All images are resized to a fixed resolution of 640 × 640 pixels and are accompanied by normalized YOLO annotation files (class_id, x_center, y_center, width, height). In total, the dataset contains 21,057 annotated bounding boxes, with an average of 3.51 objects per image. The images were generated using a synthetic data generation pipeline, where annotated waste object instances were cropped from labeled source images and composited onto real outdoor background scenes sampled from the COCO dataset. To improve realism and diversity, object placement, scale, orientation, and photometric properties were randomly varied, while quality control checks ensured annotation correctness and realistic object visibility. This dataset is intended for training, validation, and benchmarking of object detection models, particularly for challenging outdoor waste detection scenarios involving varying object scales, cluttered backgrounds, and class diversity.
Files
Steps to reproduce
The dataset was generated using a synthetic image composition pipeline. First, waste object instances were extracted from annotated source images using their bounding box annotations. Object crops with very small or visually ambiguous bounding boxes were filtered out. Outdoor background images were randomly sampled from the publicly available COCO dataset and resized or cropped to a fixed resolution of 640 × 640 pixels. For each synthetic image, multiple waste object crops were randomly selected, geometrically and photometrically augmented (random scaling, minor rotation, and brightness/contrast adjustment), and pasted onto the background image at random locations. After object placement, bounding box coordinates were recalculated and saved in normalized YOLO format (class ID, x_center, y_center, width, height). Constraints were applied to ensure realistic object scale and visibility, and invalid or overly occluded objects were discarded. This process was repeated until 6,000 synthetic images were generated. All annotation files were automatically validated, and dataset statistics were computed to verify class balance and bounding box scale diversity.
Institutions
- Chittagong University of Engineering & TechnologyChittagong, Chittagong