Citation Author(s):
Bielefeld University
Submitted by:
Christian Limberg
Last updated:
Fri, 02/28/2020 - 11:47
Data Format:
0 ratings - Please login to submit your rating.


CUPSNBOTTLES is an object data set, recorded by a mobile service robot. There are 10 object classes, each with a varying number of samples. Additionally, there is a clutter class, containing samples where the object detector failed. The data set consists of the actual object images, an HDF file with extracted visual features (via a VGG19 deep convolutional net pretrained for the imagenet competition) and a CSV file containing other information about recorded objects like their global spatial position, spatial position of the robot, a timestamp and the object position within the image


Settings of recording the data set: 

We used the robot to approach five way points to record objects that were located on different kind of furniture. On each way point, the robot did a -30 degree followed by a 60 degree turn for acquiring more object samples from different viewing angles. Objects are extracted by YOLO object detector. For simplifying our experiments, we include only object samples which have the YOLO-classes cup or bottle, however we consider these as meta categories which can be divided into different kind of e.g. bottles. Also, we exclude object samples with a YOLO-confidence with less than 20% and a distance greater than 2 meters. There are then still a share of 5\% remaining clutter images in the data set which do not contain a valid centered bottle or mug, but still exceed our YOLO confidence threshold. While the robot is moving, objects are extracted from the camera stream all the time. Since the robot moves quite fast, there are blurry objects within the data set. However, while the robot is turning at a waypoint, we make sure that all objects are detected by YOLO and there are a maximum of 3-4 frames in a row where an object is missed. The total number of frames is 515 with 2179 resulting object samples in the dataset and 121 clutter samples.


Download and extract the ZIP file containing all files. There is python code available (under 'scripts') to easily load the data set. Other programming languages should also handle .jpg, .hdf and .csv files for easy access. For easy access with python, a pickle dump file has been added. This has no extra information compared to the .csv file.