The object attributes

Participants: Aleksandrs Ecins, 'Austin Myers', Yezhou Yang, Ching Teo, 'Hui Ji', Cornelia Fermuller, Yiannis Aloimonos

The segmentation in the RGB image was used as mask to obtain the region of the depth map corresponding to the object. This depth map was then converted to a point cloud. Thus, there were two kinds of data, the RGB region and the point cloud. Using these inputs two kinds of attributes were computed: color attributes from RGB images, and shape attributes, both 2D shape attributes of form from the RGB images and 3D shape attributes from point clouds. All attributes were binary, i.e either the object has the attribute property or does not have the attribute property.

Color attributes

3D color histograms were computed from the Lab color representation. Then SVM classifier were trained for each individual color class.

2D shape attributes

We computed the following attributes on the segmentation mask.

Is long: An ellipse was fitted to the mask provided by segmentation, and the ratio of major to minor axis was used to set a threshold.

Is circular: If the ratio of major to minor axis is about the same, the object was considerer circular.

Has two parts: Along the major axis of the main ellipse, we fitted the best two ellipses. If the width of these ellipses was sufficiently different, the object was considered as consisting of two parts.

Has a handle: If one of the two ellipses obtained has a large ratio of major to minor axis, and this ellipse is longer than the other ellipse, it is considered a handle.

Has a tip: Straight edges were fitted to the object contour, and the angle at intersection computed. Small acute angles are considered as tips.

3D shape attributes

Size of object: An ellipsoid was fitted to the point cloud, the dimensions of the axes were used as size.

Is spherical: A sphere was fitted to the point cloud, and the residual was used as a threshold to decide whether the object was spherical.

Is cylindrical: A cylinder was fitted to the point cloud, and the residual of the fit was used as a threshold.

Has a flat part: If an object was classified as consisting of two parts by the 2D shape attribute method, a plane was fitted to the larger part.