ADORESet
Annotated Desktop Objects Real and Ersatz Images Dataset:
Link to the paper from the journal page:
https://rdcu.be/35Aj
A Hybrid Image Dataset Towards Bridging The Gap Between Real And Simulation Environments For Robotics
Ertugrul Bayraktar1, *
; Cihat Bora Yigit
2
; Pinar Boyraz
2 In most cases, machine vision based problems in robotics such as object detection and recognition,
object tracking and manipulation are performed using real-world or simulation images, separately. In
addition, these images routinely belong to the categories of objects namely tableware, glassware and
similar kinds of objects, which may exist on desktops, in the form of relatively small, graspable,
pushable states. In consideration of the primary purpose of robotics applications in the field of
computer vision and control in terms of imitating humans or obtaining even better results, inclusion of
irrelevant object classes in image datasets for instance wild animals, large structures, big vehicles, etc.
is infeasible which cause worse performance results. Moreover, time spent during developments in
robotics is another variable aimed to be reduced, which restricts maneuverability and diversity. In
order to optimize this period, the referred method is to utilize simulation environments, which
replicate real-world conditions as much as possible. However, simulation experiments of robotics
applications give successful results; the outcomes cannot be directly used in real-world experiments or
end-user products by means of inconsistencies between real and simulation environments. In
consequence, we propose annotated desktop objects real and ersatz images dataset (ADORESet)
(Mirror Link for Real Images, Mirror Link for Synthetic Images) that is able to accomplish these issues
with its compact form, which provides purposeful object categories with sufficient number of real and
synthetic images.
ADORESeT is composed of colored images, which has 30 classes with the dimension of 300x300 pixels.
Each class has 2500 real-world images acquired from wild web and 750 synthetic images that are
generated within Gazebo Simulation Environment (GSE). All images are stored in JPEG format. There
are 97500 images that take up about 1.8 GB space in the hard drive which makes it simpler to achieve
than ImageNet [1] (1.2 million images for 1000 classes) and MS COCO (300.000 images for 80 classes)
[2]. Although, our dataset has less total images than these two datasets, it includes more labeled
instances per category.
This hybrid dataset enables researchers to implement their algorithms both for real-world and
simulation environment conditions. Our hybrid dataset is fully-annotated and the limits of objects are
manually specified and bounding box coordinates are provided. Successor objects are also labeled to
give statistical information about the relations of the objects within dataset. For example, the relation
between monitor, keyboard and mouse can be directly obtained using this useful information.
ADORESet should be of interest to field of robotics researchers by means of its hybrid form,
compactness in terms of lightweight and relevancy to robotics applications such as detection and
recognition, grasping and dexterous manipulation of objects. 
Figure 1 Resized/labeled wild web images with instances from all categories and example images for all object categories generated in GSE.
Download Links:
ADORESet: All the content in compressed form (~1.8GB) Download
ADORESet: Real images (all classes are accessible separately) Download
ADORESet: Synthetic images (all classes are accessible separately) Download
You can find ITurk GUI at github page: https://github.com/bayraktare/ADORESet
Please Cite Us:
Bayraktar, E., Yigit, C., B., Boyraz, P. (2018). A hybrid image dataset toward bridging the gap between real and simulation environments for robotics, Machine Vision and Applications, https://doi.org/10.1007/s00138-018-0966-3.
@article{bayraktar2018a,
title={A hybrid image dataset toward bridging the gap between real and simulation environments for robotics},
author={Bayraktar, Ertugrul and Yigit, Cihat Bora and Boyraz, Pinar},
journal={Machine Vision and Applications},
doi={https://doi.org/10.1007/s00138-018-0966-3},
year={2018},
publisher={Springer-Verlag GmbH Germany} }
Object Classes:
1 Ashtray 2 Bag 3 Book 4 Bottle 5 Bowl
6 Can 7 Candlestick 8 Clock 9 CookingPot 10 Cup
11 DeskLamp 12 Eyeglass 13 ForkSpoonKnife 14 FryingPan 15 HeadWear
16 Keyboard 17 Laptop 18 Monitor 19 Mouse 20 Pen(cil)
21 PhotoFrame 22 Shoe 23 SmartPhone 24 Speaker 25 Teapot
26 Telephone 27 Vase 28 Wallet 29 WebCam 30 WristWatch
*: Corresponding Author.
1: Department of Mechatronics Engineering, Graduate School of Science Engineering and Technology,Istanbul Technical University, 34496, Maslak, Istanbul, Turkey
Tel.: +90-212-2857005
E-mail: bayraktare@itu.edu.tr
2: Department of Mechanical Engineering, Istanbul Technical University, Inonu Cd. No:65, 34437, Beyoglu, Istanbul, Turkey