In-lab Image Dataset of Foreign Objects and Anomalies in Iron Ore Conveyor Belts

This dataset contains high-speed recordings and extracted frames depicting iron ore flow on a laboratory-scale conveyor belt system, along with several classes of foreign objects (e.g., wood pieces, plastic fragments, tools) manually introduced to simulate contamination scenarios. The conveyor belt measures 35 cm in width by 1.10 m in length and is powered by an electric motor capable of speeds up to approximately 3 m/s. The overhead camera used is the onboard NVIDIA Jetson TX2 OV5693 sensor, which captured video at 120 fps and a resolution of 1280×720, using a GStreamer pipeline for direct-to-disk recording. The dataset is organized into multiple folders: Original-raw-videos: Contains the unedited MP4 files showing both normal iron ore flow and sequences with introduced foreign objects. Image-files: Includes individual frames extracted from each raw video. Subfolders are named after their corresponding source video. Image-files-manual-split: Separates frames into two categories: normal (only iron ore) and anomalous (foreign objects). Yolo-dataset-center: It provides center-cropped frames with YOLO-style labels that focus on the central region of the belt. Organized into train/test/valid splits with respective images and labels. Split-ds-normal-filtered: Offers a final curated version of the dataset, divided into normal (train/test) and anomalous frames for ease of training anomaly detection models. Scripts are included to replicate the preprocessing steps (e.g., frame extraction, YOLO-style annotations). The dataset may be used to benchmark computer vision tasks such as object detection and anomaly recognition in industrial contexts or laboratory-scale experiments. All files are unannotated by default except where explicitly labeled for demonstration purposes in the “Yolo-dataset-center” subset and partial labels for anomaly segmentation in “Split-ds-normal-filtered.” In this last folder, we only separate normal samples from anomalous ones.本数据集包含实验室规模传送带系统上铁矿石输送流的高速录像及提取帧，同时包含多类人工引入以模拟污染场景的异物（例如木块、塑料碎片、工具）。该传送带宽度为35 cm，长度为1.10 m，由最大转速约3 m/s的电动机驱动。所用顶置摄像头为搭载NVIDIA Jetson TX2 OV5693传感器的设备，以120 fps帧率、1280×720分辨率录制视频，并通过GStreamer流水线实现直接磁盘录制。 数据集按多个文件夹组织： - Original-raw-videos：包含未编辑的MP4格式原始视频文件，涵盖正常铁矿石输送流以及引入异物的录制片段。 - Image-files：包含从每段原始视频中提取的单帧图像，其子文件夹以对应源视频的名称命名。 - Image-files-manual-split：将提取的帧划分为两类：正常类（仅包含铁矿石）与异常类（包含异物）。 - Yolo-dataset-center：提供针对传送带中央区域裁剪的中心裁剪帧与YOLO风格标注文件，按训练集、测试集、验证集划分，分别包含对应图像与标注文件。 - Split-ds-normal-filtered：提供经整理的最终数据集版本，划分为正常（训练/测试集）与异常帧，以方便异常检测模型的训练。 本数据集附带可复现预处理流程的脚本（例如帧提取、YOLO风格标注生成）。本数据集可用于基准测试计算机视觉任务，例如工业场景或实验室规模实验中的目标检测与异常识别。默认情况下所有文件均未标注，仅"Yolo-dataset-center"子集以及"Split-ds-normal-filtered"中用于异常分割的部分标注文件除外。在最后一个文件夹中，仅将样本划分为正常样本与异常样本。

搜集汇总

背景与挑战

背景概述

该数据集是一个实验室规模的铁矿石传送带图像数据集，包含高速录像和提取的帧，模拟了异物（如木片、塑料碎片）污染场景，用于物体检测和异常识别研究。数据集组织为多个子集，包括原始视频、手动分割的正常与异常图像、YOLO格式标注数据，以及便于异常检测模型训练的策划版本。