Jacobs Robotics
Object Shape Category Dataset


Annotated object point clouds captured from a Kinect-like camera

Introduction
Several existing datasets are designed for evaluation of object perception tasks like detection where potential object candidates are localized in a scene, or instance recognition where the object’s identity is analyzed. The Object Shape Category Dataset (OSCD) provided here is targeted at shape-based object categorization, i.e. the task of determining that an object instance belongs in an object category which describes its general shape type. Categorization is among other tasks of interest as it contributes to the semantic understanding of objects, e.g. if an object is categorized as a cup-type, one may infer that the object is a container which could contain a liquid.

Seven different shape categories (sack, can, box, teddy bear, ball, amphora, plate) are included in the dataset with following main features:

  • Shape categories range from simple-shaped objects like balls and boxes to more complex-shaped objects like teddy bears and amphoras.

  • In addition to sensor-noise, varying self-occlusions, spatial dimensions, texture-content, objects within a category may appear deformed such as sack or teddy bears.

  • Objects of different categories can locally share similar appearing shape parts like planar parts (e.g., plates, boxes and cans) or curved parts (e.g., balls, amphoras and cans).

The dataset consists of random views on segmented objects which are captured by a Kinect-like RGB-D camera (ASUS Xtion). Subsequently, the dataset consists of single shot 2.5D RGB-D point clouds of segmented objects featuring noise and self-occlusion. As mentioned, objects of seven categories are included: for each category multiple instances were captured from multiple viewpoints each.

Technical description
In total 468 unique scans of seven categories were captured. Each category consists of multiple object instances (see preview gallery below). For each instance, multiple scans are captured from different view points. Each scan consists of a segmented object in form of a point cloud file in PCD format using the augmented 3D point type with RGB color information (cf. Point Cloud Library, pcl::PointXYZRGB). A category consists of an average of 66 scans from seven objects. For evaluation and comparison purposes, the scans are split into a training set (354 scans) and testing set (114 scans).
Each scan is only associated with a label of the corresponding category and not with a label of the individual object instance. Furthermore, no pose information or correspondence between individual scans is recorded. Thus, in this context of single-shot shape categorization, each individual scan can be interpreted as an exemplary observation of the corresponding category. This is a common practice to evaluate learned prediction models of generic object appearances like in categorization or detection tasks (compared to the case of specific object appearances in instance recognition tasks).

Download
The dataset can be downloaded in the following formats:

object_shape_category_dataset_v0_1.tar.gz object_shape_category_dataset_v0_1.zip

Citation
The Object Shape Category Dataset is part of the work presented in

Christian A. Mueller and Andreas Birk
"Visual Object Categorization Based on Hierarchical Shape Motifs Learned From Noisy Point Cloud Decompositions"
In Journal of Intelligent & Robotic Systems, May 2019

with supplementary video:

Please cite the article (DOI) if you use the dataset. Contributions to the dataset are welcome! If you have any questions regarding the dataset, please do not hesitate to contact us.
Contact: Christian A. Mueller [chr.mueller(at)jacobs-university.de]


Preview

In the follwing two different kinds of previews are provided. Please note that each category consists of multiple different object instances and scans from different viewpoints which can feature partial object observations and self-occlusions.

  1. A preview of a single scan from each category (sack, can, box, teddy bear, ball, amphora, plate) is displayed below. Additionally, for illustration purposes, object instances of each category are displayed from a viewpoint that reveals main shape properties of the respective instance.

  2. The preview below shows a random subset of unique sample scans (50% of all scans from each category) of the dataset. The full dataset can be downloaded above in the download section.


Your browser requires WebGL support in order to properly display the previews.

© 2019 Robotics Group, Jacobs University Bremen.
This website includes software developed and used by the Systems, Robotics and Vision Group, University of the Balearican Islands (srv.uib.es)
based on Bootstrap, Three.js and PapaParse.

Impressum / Imprint