The RGB-D Object Dataset provided here is for non-commercial research/educational use only.
Please cite the following paper if you use this dataset:
A Large-Scale Hierarchical Multi-View RGB-D Object Dataset
Kevin Lai, Liefeng Bo, Xiaofeng Ren, and Dieter Fox
In IEEE International Conference on Robotics and Automation (ICRA), May 2011.
Please see our paper or contact Kevin Lai (send email).
How do you extract the dataset tarballs?
Linux: Use tar, e.g. "tar xvf apple_1.tar"
Windows: To properly unpack them in Windows, you may need to extract twice using your archive program (e.g. 7zip: http://www.7-zip.org/ ). For example:
1. Open "apple_1.tar" with 7zip and extract to get "apple_1" file.
2. Open "apple_1" with 7zip and extract again to get "apple_1" folder contraining all the files.
Which objects do you leave out for category recognition experiments?
Object Recognition: For object recognition experiments on cropped images (results in figure 8 of the paper), we randomly pick one instance to leave out from each category. For the exact list of object instances left out of each random trial of our evaluation, look at the train/test splits for category recognition in the Evaluation Set.
What data is used for training and testing for the experiments in the paper?
Object Recognition (Section V of paper): We used the turntable data for both training and evaluation. Here we don't use the scene videos in the RGB-D Scenes Dataset at all.
Object Detection (Section VI of paper): We trained the detectors using the turntable data as positive examples and evaluated on the 8 video sequences in the RGB-D Scenes Dataset. During training the only videos we make use of from the RGB-D Scenes Dataset are the background videos for hard negative example mining.