introduction
What is Motion Dataset?

Motion dataset is a carefully labeled, abundant and open-source 3D model dataset for purpose of mobility analysis. Motion dataset can be used for motion part segmentation, motion attribute estimation and other related tasks. Models of the dataset are mainly from 3D warehouse and all the models are carefully labeled and orientation-aligned by students from Beijing University of Aeronautics and Astronautics (BUAA).

Download

We are glad to share this dataset to researchers around the world in computer graphics, computer vision, robotics and other relative research fields for none-commercial use. Anyone who would like to download models in mesh format (We provide FLT format and OBJs+JSON format for each model), please click this url:
http://motiondataset.zbuaa.com/download/motion_dataset/Motion Dataset v0.zip

How many object categories are in motion dataset?

In motion dataset v0, we selected 44 object categories which are common in outdoor scenes or indoor scenes and collected 2440 models in sum. The number of models from each class can be seen in taxonomy page.

What kind of label information is in motion dataset?

The most remarkable contribution of motion dataset is that it contains three kinds of carefully labeled information: "motion parts", "motion attributes" and "motion hierarchies".

Motion parts

In various categories of rigid-body models, there are some parts that can perform a reasonable movement in a certain range which are defined as "motion parts", e.g., the wheels of a car and the cap of a bottle. In our dataset, we separated the meshes of each motion part and organized them by "group nodes" in FLT models to indicate the motion parts.

Motion attributes

We summarized the most common forms of motion into three types:

We represented the motion parameters of each motion part by an axis (translation, rotation and spiral axis). Thus, motion type together with motion parameters was defined as "motion attributes". In our dataset, "degree of freedom nodes" (DOFs) in FLT models were adopted to characterize the motion attributes. What is called DOF can be simply understood as a "group node" containing a local coordinate system (LCS), according to one of the three axes of which a motion part can translate and rotate. After the location and pose of LCS are set correctly, it will be easy to control the motion part to perform a reasonable movement.

Motion hierarchies

Under quite a few circumstances, the movement of a motion part is not considered to be independent. For instance, when the front frame of a bicycle is turning, its front wheel will be driven to move simultaneously. As a result, the movement of the front wheel is actually the combination of both turning and rolling. Situations being similar to this are resulted by the “motion hierarchies”of models. We applied the tree structure of FLT models to illustrate the motion hierarchies: the motion parts controlling other motion parts were set as a parent node, while the motion parts being implicated were set as children nodes.