Floating Waste and Aquatic Vegetation Image Dataset from Turbid Rivers

Description

This dataset contains annotated RGB images of floating waste and river vegetation collected from turbid river environments in Banjarmasin, South Kalimantan, Indonesia. The dataset was compiled from two sources: still photographs and image frames extracted from video recordings. Video frames were extracted at a rate of approximately one frame per second. The collected images were manually screened to remove images that did not contain relevant objects or were unsuitable for annotation. The final dataset consists of 1,577 unique images derived from 24 video sources and still photographs. Each image was manually annotated using bounding boxes for two object categories: floating waste and river vegetation. The annotations are provided in YOLO format, with class identifiers and normalized bounding-box coordinates. The dataset contains a total of 5,792 annotated object instances, comprising 3,718 floating waste instances and 2,074 river vegetation instances. For model development and evaluation, the dataset is divided into training, validation, and testing subsets using an approximately 70:20:10 ratio, containing 1,104 training images, 315 validation images, and 158 testing images. To reduce the risk of data leakage, images originating from the same video source were kept within the same dataset subset during partitioning. Thus, frames extracted from a single video source were not distributed across different training, validation, and testing subsets. The dataset is intended to support research in computer vision, object detection, environmental monitoring, floating waste detection, river monitoring, and related applications. The dataset can also be used for benchmarking and evaluation of object detection models under challenging visual conditions, including varying water turbidity, illumination, object distribution, and complex river backgrounds. The dataset is provided together with the image files, YOLO annotation files, and dataset configuration information to facilitate reproducible object detection experiments.

Steps to reproduce

1. Collect still photographs and video recordings from turbid river environments in Banjarmasin, South Kalimantan, Indonesia. 2. Extract image frames from the recorded videos at approximately one frame per second. 3. Screen the collected images and remove images that do not contain relevant objects or are unsuitable for annotation. 4. Manually annotate the remaining images with bounding boxes for floating waste and river vegetation using YOLO annotation format. 5. Organize the annotated images into training, validation, and testing subsets using a group-aware partitioning strategy, ensuring that frames originating from the same video source remain within the same subset.

Institutions

• Lambung Mangkurat University

  South Kalimantan, Banjarmasin

Categories

Licence