Construction and Demolition Waste Object Detection Dataset

The Construction and Demolition Waste (CDW) dataset was created to facilitate the development and benchmarking of object detection models responsible for the localization and classification of three common object classes of CDW, namely concrete, brick and tile in CDW sorting facilities, under working conditions. For the development of the training and testing datasets, samples belonging to these three object classes were extracted from manually sorted piles of CDW from a recycling facility in Cyprus and digital images of the material were recorded in a controlled environment on the conveyor belt of a full-scale model sorting platform with the use of a colored (RGB) camera. A total of 550 .JPG mages were recorded (containing approximately 4 samples of each object class) at a resolution of 1920 x 1200 x 3, and a total of approximately 6600 samples of all object categories were attained. We emphasized on the complexity of CDW in working conditions by capturing the heterogeneity, surface contamination, irregularity as well as the adherence and stacking of samples in two independent testing datasets. The first testing dataset (testing_set_1) represents an idealized case of CDW placement where the samples are sparsely spaced on the conveyor belt and serves as a baseline indicator of model performance. The second dataset (testing_set_2) contains heavily stacked and adhered samples of CDW so as to give a more representative evaluation of the models under working conditions. Accordingly, we split the samples in three sets with the following distribution: - Training set: 4230 samples are used for training the object detectors - Testing_set_1 : 1727 samples are used for testing the object detectors on an idealized case of CDW placement - Testing_set_2: 596 samples are used for testing the object detectors on heavily stacked and adhered cases To facilitate implementation, in tandem with the provided images, the dataset contains carefully annotated ground-truth bounding boxes in .xml format.本建筑与拆除垃圾（Construction and Demolition Waste，简称CDW）数据集旨在助力CDW分拣场景下三类常见CDW目标的定位与分类任务的目标检测模型开发与性能基准测试，这三类目标分别为混凝土、砖块与瓦片。 为构建训练与测试数据集，研究人员从塞浦路斯一家回收厂的人工分拣CDW堆中提取了上述三类目标样本，并在全尺寸模拟分拣平台的传送带场景下，于可控环境中使用彩色（RGB）相机采集了物料的数字图像。本次采集共获得550张分辨率为1920×1200×3的JPG格式图像，单张图像平均包含约4个目标样本，最终累计得到约6600个全类别目标样本。 为还原作业场景下CDW的真实复杂性，我们在两个独立测试数据集的采集过程中，刻意纳入了样本异质性、表面污染、形态不规则性，以及样本黏附与堆叠等特征。第一个测试数据集（testing_set_1）为理想化场景：样本在传送带上稀疏排布，可作为模型性能的基准评估指标。第二个测试数据集（testing_set_2）则包含大量堆叠且相互黏附的CDW样本，用于更贴合实际作业场景的模型性能评估。 据此，我们将全部样本划分为以下三个子集： - 训练集：共4230个样本，用于目标检测模型的训练 - 测试集1（testing_set_1）：共1727个样本，用于在CDW样本稀疏排布的理想化场景下测试模型性能 - 测试集2（testing_set_2）：共596个样本，用于在样本高度堆叠黏附的场景下测试模型性能 为便于模型落地使用，本数据集除提供原始图像外，还附带了经精准标注的XML格式真实边界框（ground-truth bounding boxes）标注文件。