RGB-D Part Affordance Dataset

To cut an apple or hammer a nail, robots need to not just recognize the tool's name, but they must identify the tool's parts and their functions, or affordances. Recognizing the affordances of objects and their parts could allow robots to generalize to large and diverse object categories.

The RGB-D Part Affordance Dataset contains RGB-D images and ground-truth affordance labels for 105 kitchen, workshop and garden tools, and 3 cluttered scenes. The dataset provides a large and diverse collection of everyday tools, where tools from different categories share the same affordance. This enables a variety of experimental settings such as zero-shot learning. There are seven affordances associated with the surfaces of tool parts: grasp, cut, scoop, contain, pound, support and wrap-grasp. In addition to pixel-wise labels of the most likely affordance of each part, we also provide an ordered ranking of multiple affordances.

If you use this dataset, please cite the following paper:

Affordance Detection of Tool Parts from Geometric Features,
Austin Myers, Ching L. Teo, Cornelia Fermüller, Yiannis Aloimonos.
International Conference on Robotics and Automation (ICRA). 2015. [project page]


Questions? Please contact amyers "at" umiacs "dot" umd "dot" edu