Indoor Object Detection Dataset

The indoor object detection dataset is divided into three parts: the training set (94%), validation set (4%), and test set (2%), with 12,012 images for training, 490 for validation, and 245 for testing. The dataset contains a total of 12,747 images. The dataset is organized into seven classes, which are fire extinguisher, shelf, door, table, human, chair, and bin. The dataset class distribution for 7 classes: 2,693 fire extinguishers, 1,384 shelves, 8,530 doors, 3,959 tables, 5,010 humans, 9,687 chairs, and 3,475 bins. Moreover, in the training split, the chair class is most prevalent with 9,090 instances, followed by doors (8,050), humans (4,711), tables (3,731), bins (3,275), fire‑extinguishers (2,547) and shelves (1,316). The validation set shows a similar hierarchy—chairs (429) leading, then doors (305), humans (216), tables (161), bins (135), fire‑extinguishers (97) and shelves (48). In the test split, doors edge ahead with 175 instances, while chairs follow closely at 168, then humans (83), tables (67), bins (65), fire‑extinguishers (49) and shelves (20). In addition, preprocessing steps included auto-orientation and resizing all images to 640×640. To improve generalization for real-world applications, we applied data augmentation techniques, including horizontal and vertical flipping, 90-degree rotations (clockwise, counterclockwise, and upside down), random rotations within -15° to +15°, shearing within ±10° horizontally and vertically, and brightness adjustments between -15% and +15%. Additionally, this annotated, preprocessed, and augmented dataset enhances object detection accuracy in indoor scenes.

本室内目标检测（object detection）数据集分为训练集（占比94%）、验证集（占比4%）与测试集（占比2%）三个子集，其中训练集包含12012张图像，验证集490张，测试集245张，数据集总图像量为12747张。该数据集共涵盖7个类别，分别为灭火器（fire extinguisher）、货架（shelf）、门（door）、桌子（table）、人体（human）、椅子（chair）与垃圾桶（bin）。各分类的总标注数量分布如下：灭火器2693个，货架1384个，门8530个，桌子3959个，人体5010个，椅子9687个，垃圾桶3475个。在训练子集内，椅子类的实例数量最多，达9090个，其次为门（8050个）、人体（4711个）、桌子（3731个）、垃圾桶（3275个）、灭火器（2547个）与货架（1316个）。验证集的类别分布呈现相似的层级特征：椅子（429个）居首，随后依次为门（305个）、人体（216个）、桌子（161个）、垃圾桶（135个）、灭火器（97个）与货架（48个）。测试子集则以门类略占优势，包含175个实例，紧随其后的是椅子（168个），后续依次为人体（83个）、桌子（67个）、垃圾桶（65个）、灭火器（49个）与货架（20个）。此外，数据集的预处理步骤包含自动方向校正与将所有图像统一调整至640×640分辨率。为提升模型在实际场景中的泛化能力，我们采用了多种数据增强（data augmentation）技术，包括水平与垂直翻转、90度旋转（顺时针、逆时针与上下翻转）、-15°至+15°范围内的随机旋转、水平与垂直方向±10°的剪切变换，以及亮度调整范围为-15%至+15%的亮度校正。该经过标注、预处理与数据增强处理的数据集，可有效提升室内场景下的目标检测精度。