DSLR and Smartphone Camera-Based Plastic Waste Detection Dataset from Cox’s Bazar Sea Beach, Bangladesh

Description

The Cox’s Bazar Beach Plastic Waste Detection Dataset (CoxPWD 2025) is an object detection dataset of 4,195 RGB images collected from Cox’s Bazar, Bangladesh. Images were captured at Kolatoli Beach, Sugandha Beach and Laboni Beach using an iPhone 15 Pro Max, Google Pixel 6 and Canon 650D under sunny, cloudy and rainy conditions during morning, noon and afternoon, with camera heights of approximately 1 m and 6 m and viewpoints of approximately 15° and 45° from the left and right to capture realistic variation in scale, occlusion, lighting and background. All images are stored as 640×640 JPEG files and annotated for bounding box object detection in COCO format with 14 plastic waste classes which are packet, polythene, cup, spoon, straw, rope, bag, bottle, bottle_cap, net, sunglass, toy, fishing_item and others. The dataset is split into 80 percent training (3,367 images), 10 percent validation (414 images) and 10 percent test (414 images) with a separate _annotations.coco.json file for each subset, making it directly usable for YOLOv8 and other deep learning models for coastal plastic waste monitoring and benchmarking.

Steps to reproduce

To reproduce the dataset preparation pipeline: 1. Data collection Conduct ground-level surveys along Kolatoli, Sugandha, and Laboni beaches in Cox’s Bazar. Using an iPhone 15 Pro Max, Google Pixel 6, and Canon 650D, photograph visible plastic waste in portrait and landscape modes from approximately 15° and 45° left/right viewpoints and at about 1 m and 6 m camera heights, under sunny, cloudy, and rainy conditions during morning, noon, and afternoon. 2. Curation Transfer all images to a computer and remove blurry, duplicate, empty, or ambiguous images, retaining only frames where plastic waste is clearly visible. 3. Annotation Import the curated images into Roboflow. Define the 14 plastic waste classes packet, polythene, cup, spoon, straw, rope, bag, bottle, bottle_cap, net, sunglass, toy, fishing_item and others and manually draw bounding boxes around each instance, followed by quality control correction. 4. Pre-processing and split Resize all images to 640×640 RGB without applying data augmentation, then create training, validation, and test splits in an 80/10/10 ratio. 5. Export and packaging Export the dataset in COCO object-detection format, generating _annotations.coco.json files inside train/, valid/, and test/ folders, organized as: plastic/train/*.jpg and _annotations.coco.json plastic/valid/*.jpg and _annotations.coco.json plastic/test/*.jpg and _annotations.coco.json

Institutions

• East West University

  Dhaka District, Dhaka

Categories

Licence