TrashNet++: A Real-Time Multi-Class Waste Recognition Dataset for Smart City and IoT Applications
Description
The WasteVision dataset is a high-quality image dataset collected from the real world with the aim of assisting in object detection and classification for non-biodegradable solid waste in cities. It contains 2,600 high-resolution RGB images (with 8,320 objects) annotated and labeled into seven classes of wastes. The collection aims to provide strong support for building computer vision models in smart city applications, specifically in low-resource settings. The photographs were shot from two socioeconomically and environmentally distinct urban zones of Bangladesh: - Khulna Sadar (22.8456° N Latitude, 89.5403° E Longitude) - Dhaka Pallabi (23.8261° N Latitude, 90.3654° E Longitude) To simulate the uncertainty of actual sensors, the images in the collection were captured by three mid-range camera devices: - A Canon EOS 1300D digital single-lens reflex - An iPhone 12 Pro with a 12-megapixel camera - A Samsung Galaxy A51, a second 12-megapixel midrange smartphone Each image is saved in JPEG (.jpg) format and resized to 640×640 pixels evenly with bicubic interpolation. The resolution was chosen to enable standardizing training pipelines and interoperability with famous object detection models such as YOLOv5 and YOLOv8. The data is marked up in YOLO bounding box format as (normalized coordinates x-center, y-center, width, height) and class IDs for each object. The annotations were all done manually via Roboflow with great accuracy and then also verified by multiple annotators to maintain consistency. For each image, there is a corresponding .txt file with bounding box information, which is saved in separate label directories reflecting the YOLO directory layout. The data is in three structured subsets: - Training set has 1,825 images and 5,913 waste object annotations. - Validation set has 515 images and 1,634 annotations. - Test set has 260 images and 773 annotations. This separation makes it easy to have plug-and-play compatibility with deep learning workflows, especially YOLO-based pipelines, with a layout like images/train/, images/val/, images/test/ and corresponding labels/ directories. The information includes seven impact-dense waste types according to urban frequency, health risk, and ecological importance: - Plastic: Includes plastic boxes, packaging, and bottles, which dominate urban waste streams. - Metal: Includes metal rods, sheets, and boxes; typically industrial and construction. - Glass: Includes crushed glass and bottles made of glass, typically in bins or mixed with other trash. - Can: Typically aluminum beverage cans discarded in commercial districts or near public events. - Cable: Refers to electric cables and wires, which is the growing problem of technology waste. - E-waste: Comprises dumped chargers, circuit boards, and faulty electronics, with high possibilities of toxicity. - Medical Waste: Comprises used masks, gloves, syringes, and medicine vials, especially from hospitals or roadside dumps following the pandemic.
Files
Institutions
- Daffodil International University