Accuracy improvement of functional attribute recognition by dense CRF considering object shape

In this paper, we propose a method to recognize functional attributes of everyday objects for vision system of partner robots. On the related research, there is a method to optimize recognition result with dense (fully connected) conditional random field (CRF), which uses the estimation result of functional attribute for each pixel. However, since this method is optimized from RGB data, it is not able to sufficiently consider the shape of object, which has a relationship with the function attribute. In the proposed method, the recognition accuracy of functional attributes is improved by considering the object shape with the dense CRF describing the three-dimensional positional relationship. As a result of the experiment, the recognition rate of the proposed method is 77.0%, which is 3.8% higher than the related method. In addition, we confirmed that the processing speed is high as a side effect by reducing processing cost by oversegmentation of input data and using high-speed identification by Random Forests. The mean processing speed per an object was 109 ms in the proposed method.