Walking-Pose with Exoskeleton

We provide an RGB-D dataset with fine annotation on the detection of joint points during human walking, including 3588 pairs of RGB images and corresponding depth images. These RGB-D images are captured from persons walking freely and wearing the exoskeleton in different scenes.
The training set contains 2,870 pairs of RGB-D images, and the validation set contains 718 pairs of RGB-D images. We use Realsense D435i depth camera to collect images from actual scenes at Beihang University, and participants include six males and four females. These images are padded and resized to 512 x 512 to simplify the data loading process.
Each annotation contains eight human joint points. The annotation form of the dataset is organized as follows:
x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 x6 y6 x7 y7 x8 y8 /n
...
Where,
x1 and y1 — left shoulder joint.
x2 and y2 — right shoulder joint
x3 and y3 — left hip joint
x4 and y4 — right hip joint
x5 and y5 — left knee joint
x6 and y6 — right knee joint
x7 and y7 — left ankle joint
x8 and y8 — right ankle joint
The label of an image is stored in a text file and associated by the file name.

我们构建了一款面向人体行走关节点检测任务的精细化标注RGB-D数据集，涵盖3588组RGB图像与对应深度图像对。所有RGB-D图像均采集自不同场景下自由行走且穿戴外骨骼的受试者。
训练集包含2870组RGB-D图像对，验证集包含718组RGB-D图像对。本数据集采用RealSense D435i深度相机，在北京航空航天大学的真实场景中采集图像，受试者共计10名，其中男性6名、女性4名。为简化数据加载流程，所有图像均经过填充与缩放处理，统一调整至512×512分辨率。
每条标注包含8个人体关节点信息，数据集的标注格式如下：
x1 y1 x2 y2 x3 y3 x4 y4 x5 y5 x6 y6 x7 y7 x8 y8
其中：
x1与y1对应左肩关节点；
x2与y2对应右肩关节点；
x3与y3对应左髋关节点；
x4与y4对应右髋关节点；
x5与y5对应左膝关节点；
x6与y6对应右膝关节点；
x7与y7对应左踝关节点；
x8与y8对应右踝关节点。
单张图像的标注标签存储于文本文件中，并通过文件名与对应图像建立关联。