# Granular stockpile volume dataset

## TL;DR

This paper introduces a comprehensive UAS dataset for measuring granular stockpile volumes using vision-based methods, supporting research in automation and accuracy.

## Contribution

The paper presents a novel annotated UAS dataset for granular stockpile volume measurement with varied data collection parameters.

## Key findings

- The dataset includes 1521 images and 3D models of 47 stockpiles with volumes ranging from 51 to 3000 m³.
- Point clouds were annotated in PLY and XYZ formats, enabling 3D deep learning for object detection.
- The dataset supports cross-sectional analysis of stockpiles under varying environmental and flight conditions.

## Abstract

Unmanned Aerial Systems (UAS) applications are growing for vision-based volume measurement to enhance accuracy, efficiency, and automation. Despite the growing applications of UAS, no comprehensive dataset is currently available for researchers to determine the effects of visual data collection parameters such as camera angles, image overlaps, and flight patterns, on the outcomes. These outcomes consist of but are not limited to the number of images, the density of point clouds, and the quality of 3D models. This study introduces an annotated UAS dataset to allow researchers and practitioners to use vision-based UAS data for accurate measurement of granular stockpiles. Data were collected from stockpiles with irregular shapes in Grand Forks, ND, USA using UAS. The dataset includes 1521 images captured under varying weather conditions, stockpile sizes, camera angles, flight patterns, flight heights, and image overlaps. This study investigated 47 stockpiles across two distinct sites, including sand and gravel materials. Using Pix4D photogrammetry, 3D models were generated, with individual stockpile volumes ranging from 51 m³ to 3000 m³. Data was collected during multiple surveys; however, stockpiles were not individually tracked across time, so the dataset should be regarded as cross-sectional rather than strictly longitudinal. Stockpile volumes in one of the sites changed overtime during the data collection. The dataset was enriched with annotated 3D points identifying not only stockpiles, but irrelevant objects, such as trees, vehicles, and roads. The point clouds generated from these models were annotated in PLY and XYZ formats, creating a unique 3D point dataset with corresponding 2D images. This dataset is well-suited for the development of autonomous detection and measurements of objects using 3D deep learning models for object detection.