River Obstacle Segmentation En-route By USV Dataset (ROSEBUD)

<p dir="auto">Supervised Semantic Image Segmentation has proven to be a robust and effective method to obtain dense feature and location information from RGB images. Such methods rely on neural network architectures being trained on ground truth data to obtain network weights and biases that accurately provide feature information for RGB input images. Many datasets exist of ground and aerial scenes that can be used for training of neural networks built to aid in navigation of unmanned aerial and ground vehicles. Fewer datasets that aid in the recognition of water for the determination of navigable areas for Unmanned Surface Vehicles (USV) exist and fewer yet exist that can aid in the segmentation of fluvial (river) scenes. Such datasets are needed to train segmentation agents in properly interpreting visual data from complex fluvial scenes to allow for efficient autonomous navigation of  USVs (boats) down rivers and creeks.</p>

<p dir="auto">To help solve the data sparsity issue this repository contains images and ground truth masks of the River Obstacle Segmentation En-route By USV Dataset (ROSEBUD) Fluvial Dataset. The dataset contains annotated images for training supervised networks on the task of recognizing and segmenting fluvial scnenes as well as water. In total there are 549 images in the data obtained on Sugar Creek and the Wabash River in the US state of Indiana during July and September of 2021 respectively. Annotated sematic masks are provided as both binary (water and non-water) as well as multi-class (water, sky, foliage, boats/people, debrise/obstacles, river bank/shore, and bridge) problems. The masks are also provided as RGB (3-channel) as well as greyscale (1-channel) in both standard and high definition formats. Mask descriptions, a structure of the files, and class RGB/greyscale values are recorded in the readme.</p>