We provide 5 different datasets for evaluating 3D keypoint detectors’ performance. All datasets come with groundtruth.

• Retrieval: 18 scenes (at 3 different noise levels) and 6 models [SCENES] [MODELS]
• Random Views: 36 scenes (at 3 different noise levels) and 6 models (Note: the models for this dataset are the same as for the Retrieval one) [SCENES] [MODELS]
• Laser Scanner: 50 scenes and 5 models [ SCENES & MODELS ( please see bottom note *)]
• Space Time: 12 scenes and 6 models [SCENES & MODELS]
• Kinect: 17 scenes and 27 models [SCENES & MODELS]

From top to bottom row: Random Views, Laser Scanner, Space Time, Kinect

*NOTE: The dataset “Laser Scanner” is here redistributed by courtesy of Prof. Ajmal Mian, who is the owner of the dataset and of the copyright linked to it. A link to the original repository of the dataset can be found here. We wish to remind that this dataset is provided without any warranty and for research purposes only. If this dataset is used, the following two papers are to be cited:

• Ajmal Mian, M. Bennamoun and R. Owens, “3D Model-based Object Recognition and Segmentation in Cluttered Scenes”, IEEE Trans. on PAMI, vol. 28(10), pp. 1584–1601, 2006.
• Ajmal Mian, M. Bennamoun and R. Owens, “On the Repeatability and Quality of Keypoints for Local Feature-based 3D Object Retrieval from Cluttered Scenes”, IJCV, vol. 89(2), 348–361, 2010.