Kinect RGB-D Image Dataset

I published a dataset of Kinect RGB-D images. This dataset also provides ARGs and graph templates used in our research, which fits the requirements of learning graph matching. This dataset consists of RGB-D images containing about 1200 objects. These RGB-D images are collected in different environments, indoors and outdoors. There are ten large categories, such as basket, bucket, drink box, bicycle, scanner, fridge, notebook PC, sprayer, dustpan, and platform lorry. Each category has a large number of objects. The RGB-D image is in the size of 640×480. Objects inside a category usually have different textures, and they are placed in complex environments with different translations and rotations. Moreover, objects within some categories have large intra-category variations in size and local structure. For example, bicycles for men have beams, while those for women do not. Small bicycles are usually with simpler structures, and compared to other parts, the wheel radius changes most in size among different bicycles.

Citation

Quanshi Zhang, Xuan Song, Xiaowei Shao, Huijing Zhao, and Ryosuke Shibasaki, “Category Modeling from just a Single Labeling: Use Depth Information to Guide the Learning of 2D Models” in Proc. of IEEE International Conference on Computer Vision and Pattern Rcognition (CVPR), 2013.

We also share graphs and initial graph templates used in “Q. Zhang, X. Song, X. Shao, H. Zhao, R. Shibasaki, Learning Graph Matching for Category Modeling from Cluttered Scenes, ICCV 2013″.

You may download the dataset with graphs and templates (dataset1, and dataset2 (with graphs)). If you have questions, please contact Dr. Quanshi Zhang.