**7. Experimental Results**

Figure 9 illustrates the 3D objects used in this study. Twenty-four object models belonging to eight classes are selected from a popular dataset [25]. We have applied the proposed framework to acquire sequential tactile data by blind contour following as well as using contour following paths guided by visually interesting points and the obtained sequences are classified using the previously explained approaches.

Since CNN, like any other deep learning solution, requires a large data set for training; we have collected a total of 22,800 sequences, from which 20% (4560) is used for testing and the obtained accuracy is reported in Table 1. The process of tactile data acquisition is simulated using the MATLAB programming platform and its statistics and machine learning toolbox are used for training and testing Convolutional Neural Networks.

In the case of the approach in Section 6.2, the acquired data set is first standardized using z-score before being fed into SVM and kNN. After splitting the data into 80:20 samples for developing the classifier and testing it, the conventional classifiers are trained and validated using 5-fold cross-validation. The kNN classifier takes benefit from the Euclidean distance metric to determine nearest neighbors and assigns the label of winning vote among the 10 nearest ones to test data. The support vector machine employs an RBF kernel function with *γ* = 1, = 0.5, and *C* = 1 hyperparameters.

**Figure 9.** Objects used for experiments.

The accuracy values for the 20% of the data kept out for testing, which includes 4560 test sequences, are reported in Table 1 for the three classifiers. Confusion matrices are also provided in Figure 10, in which the eight object classes are numbered as class one to eight in the same order in which they are presented in Figure 8.

The accuracy values confirm that the use of visually interesting points to determine object contours has a positive impact on classification accuracy for all classifiers and in the case of Convolutional Neural Networks the accuracy is improved by 15.68%. Furthermore, CNNs show a good capability in extracting and learning features from tactile data. It is worth mentioning that the random guess in our experiments is <sup>1</sup> <sup>8</sup> or 12.5%, and the best performance achieved in this work is 98.97%, i.e., 86.47% above the random guess.




**Figure 10.** Confusion matrices.

According to the confusion matrices, the visual data helps making a cleaner discrimination among some object classes, so the confusion occurs only among the classes with more tactile similarities. This can be explained by the fact that visually interesting points lead to the selection of contours which are more informative about the object characteristics. For example, a round or oval shape contour can be followed on almost all objects if we blindly follow a local path around object extremities, while visual data guides the process by selection of different contours simplifying the object recognition process.
