Object recognition from images that is one of the crucial and demanding areas of the real world applications has become a popular investigation issue with the development of low cost depth sensing cameras like Microsoft Kinect and methods such as Deep Learning. The use of Convolutional Neural Networks (CNN), a well-known Deep Learning technique on RGB-D images that provide color and depth information, has significantly increased the correct recognition of objects. Despite their success, CNNs have disadvantages such as they can not tolerate viewpoint variations and loses the spatial information among the features. To overcome the disadvantages of CNNs, Capsule Networks has been offered by Hinton. A Capsule is a small group of neurons that represent the various features of a particular entity that are present in the image. While the length of the activity vector of a capsule represents the probability that an object is present in the image, the orientation of it represents the instantiation parameters such as pose, velocity, albedo etc. The main aim of this study is to evaluate the object recognition performance of Capsule Networks on RGB-D images by using a subset of Washington RGB-D dataset. For this purpose, two different Capsule networks are designed separately for RGB and Depth images. Grayscale depth images are colorized and all the images are resized. The object recognition performances of the Capsule networks are evaluated on the preprocessed images and compared with the CNNs. According to the results, the performance results of the Capsule networks are promising.
Anahtar Kelimeler: Capsule Network, RGB-D Data, Object Recognition
|
