The PartNet Symmetry Hierarchy Dataset (PartNet-Symh)

Introduction

The PartNet-Symh dataset augments the PartNet dataset by adding a recursive hierarchical organization of fine-grained parts for each shape. The PartNet dataset was originally proposed in the paper "PartNet: A Large-scale Benchmark for Fine-grained and Hierarchical Part-level 3D Object Understanding". The hierarchical organization follows the symmetry hierarchy defined in [Wang et al. 2011]. The symmetry hierarchies were used to train the PartNet model proposed in our paper "PartNet: A Recursive Part Decomposition Network for Fine-grained and Hierarchical Shape Segmentation". In general, PartNet-Symh can be used to train any model for encoding/decoding part-based structures based on Recursive Neural Networks, e.g., GRASS [Li et al. 2017].

Basic Information

The dataset contains 22369 3D shapes covering 24 shape categories. See Table 1 for the statistics of the dataset.

category	Bag	Bed	Bottle	Bowl	Chair	Clock	Display	Door	Faucet	Hat	Keyboard	Knife	Lamp	Laptop	Microwave	Mug	Refrigerator	Scissors	Storage	Table	TrashCan	Vase	Dishwasher	Earphone
# shapes	154	147	511	102	6201	426	329	198	826	251	109	486	2603	92	81	231	164	112	2598	5868	296	180	135	269
# parts	343	3534	1432	213	38858	1152	1174	586	3986	588	5690	1571	12567	270	346	529	671	394	35507	30408	2580	447	928	1193
max.# parts per shape	4	89	5	3	30	8	5	9	18	3	64	5	128	3	8	4	8	5	101	50	43	4	8	8
min.# parts per shape	2	4	2	2	2	2	2	2	2	2	13	2	2	2	3	2	2	2	2	2	2	2	2	2

Table 1. Statistics of the PartNet-Symh dataset.

Explaining the Dataset with an Example

Let us use a chair as an example to illustrate how our data is organized. We first show a figure to illustrate how to represent a part-based model with a symmetry hierarchy. We then explain the details of data organization.

1. Symmetry hierarchy

Figure 1. A chair model is represented with a symmetry hierarchy which is a top-down recursive decomposition into its constituent parts.

As shown in Figure 1, a chair model is represented with a recursive symmetry hierarchy. Each leaf node represents a part. There are three types of nodes in the hierarchy: Type 0 -- Leaf nodes (e.g. node 2), Type 1 -- Adjacency nodes (e.g. node 11, indicating the proximity relations between two adjacent parts), and Type 2 -- Symmetry nodes (e.g. node 9 or node 12, which represents either a reflectional or a rotational symmetry relations of multiple parts). A symmetry node stores the parameters (e.g., reflection axis) of the corresponding symmetry. Please refer to [Wang et al. 2011] and [Li et al. 2017] for more detailed definition of symmetry hierarchy.

We ensure all shapes belonging to the same shape category share the same high-level structure of symmetry hierarchy. This means that those shapes have consistency in the top few levels of their symmetry hierarchies [van Kaick et al. 2013]. These levels generally correspond to semantically meaningful major parts. For example, a chair model is composed of a back, a seat, a leg and an armrest, and the symmetry hierarchies are consistent at the level of these parts across all chairs.

2. Data organization

There are seven folders for each model.

A. The ops folder

Each mat file in this folder stores the corresponding types of the nodes of a symmetry hierarchy. Taking the symmetry hierarchy in Figure 1 (b) for example, Table 2 gives the node types in a depth-first traversing order with vertex postorderings.

node	node 7	node 3	node 4	node 11	node 12	node 13	node 6	node 1	node 2	node 8	node 9	node 10	node 14	node 5	node 15
type	0	0	0	1	2	1	0	0	0	1	2	1	1	0	1

Table 2. Node type (0 -- leaf node, 1 -- adjacency node, and 2 -- symmetry node) of the nodes in Figure 1 (b).

B. The part mesh indices folder

The mat file under this folder stores the part mesh indices corresponding to the leaf nodes of a symmetry hierarchy. Taking the symmetry hierarchy in Figure 1 (b) for example, Table 3 gives the part mesh indices of leaf nodes in the same order as above.

leaf node	node 7	node 3	node 4	node 6	node 1	node 2	node 5
part mesh indices	6	5	4	7	16	9	1

Table 3. part mesh indices for leaf nodes.

For example, you can find the sixth part mesh for node 7 in 'objs' class from result_after_merging.json file for this shape. Note that this json file can be found in dataset from [Mo et al 2019]. There are many 'objs' classes in this json file, please make sure only use the last 'objs' class which contains names of all parts.

Shape name mapping from [Mo et al 2019] to ours: a shape_name field of each mat file indicates the original shape id from [Mo et al 2019].

C. The boxes folder

The mat file under this folder stores the parameters of the part bounding boxes corresponding to the leaf nodes of a symmetry hierarchy.

D. The labels folder

The mat file under this folder stores the semantic label for each leaf node. Table 4 gives the node labels of a chair model, node 5 is the back part of the chair, labeled as 0. node 7 is the seat, labeled as 1. node 1, node 2 and node 6 are the leg parts labeled as 2. node 3 and node 4 represent the armrests labeled as 3.

node	node 7	node 3	node 4	node 6	node 1	node 2	node 5
label	1	3	3	2	2	2	0

Table 4. Node labels (0 -- back, 1 -- seat, 2 -- leg, and 3 -- armrest).

E. The syms folder

The mat file in the syms folder stores the symmetry parameters for each symmetry node. The symmetry parameters defined for each symmetry type can be found in [Li et al. 2017].

F. The models folder

The models folder stores the 3D mesh models in .obj format.

G. The obbs folder

The obbs folder stores the whole shape obb for each model, which contains original part obb, adjacent part relations and symmetric parameters.

Downloading

The dataset can be downloaded from here.

Read data

Code for reading symmetry hierarchies and part bounding boxes can be found here.

Code for sampling point cloud from part mesh, reading symmetry hierarchies and part point clouds can be found here.

Reference

[Wang et al. 2011] Yanzhen Wang, Kai Xu, Jun Li, Hao Zhang, Ariel Shamir, Ligang Liu, Zhi-Quan Cheng, and Yueshan Xiong, "Symmetry Hierarchy of Man-Made Objects", Computer Graphics Forum (Special Issue of Eurographics 2011), 30(2): 287-296.

[van Kaick et al. 2013] Oliver van Kaick, Kai Xu, Hao Zhang, Yanzhen Wang, Shuyang Sun, Ariel Shamir and Daniel Cohen-Or, "Co-Hierarchical Analysis of Shape Structures", ACM Transactions on Graphics (SIGGRAPH 2013), 32(4).

[Li et al. 2017] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang and Leonidas Guibas, "GRASS: Generative Recursive Autoencoders for Shape Structures", ACM Transactions on Graphics (SIGGRAPH 2017), 36(4).

Citation

If you use this dataset, please cite the following papers.

@InProceedings{Yu_2019_CVPR,
    title = {{PartNet}: A Recursive Part Decomposition Network for Fine-grained and Hierarchical Shape Segmentation},
    author = {Fenggen Yu and Kun Liu and Yan Zhang and Chenyang Zhu and Kai Xu},
    booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    pages = {to appear},
    month = {June},
    year = {2019}
}

@InProceedings{Mo_2019_CVPR,
    author = {Mo, Kaichun and Zhu, Shilin and Chang, Angel X. and Yi, Li and Tripathi, Subarna and Guibas, Leonidas J. and Su, Hao},
    title = {{PartNet}: A Large-Scale Benchmark for Fine-Grained and Hierarchical Part-Level {3D} Object Understanding},
    booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2019}
}