Abstract:

10 categories, 40 objects acquired in human mode (see definition above) for the training phase. Data for the testing phase have been collected both in human and robot mode. The acquisition size is 640x480 and subsequently cropped to the bounding box of the object according to the kinematics or motion cue. The bounding box is 160x160 in humand mode and 320x320 in robot mode.

For each object we provide 200 training samples. Each category is trained with 3 objects (600 examples per category).

The following tests are provided in order to assess performance: Demonstrator ? During the test we change the demonstrator, the objects are known instances; Categorization ? We test new instances of objects; Robot ? We let the robot to grasp the objects and to recognize their category. For this test we use both known and unknown instances; Background ? We select 10 images per category where the classifiers perform 99% of accuracy. We provide the segmentation mask of the objects and we test if the classifier recognizes the object or the background.