Applicability Of Deep Learned Vs Traditional Features For Depth Based Classification

In robotic applications, highly specific objects such as industrial parts, for example, often need to be recognized. In these cases methods can't rely on the online availability of large labeled training data sets or pre-trained models. This is especially true for depth data, thus making it challenging for deep learning (DL) approaches. Therefore, this work analyzes the performance of various traditional (global or part-based) and DL features on a restricted depth data set, depending on the tasks complexity. While the sample size is small, we can conclude that pre-trained DL descriptors are the most descriptive, but not by a statistically significant margin and therefore part-based descriptors are still a viable option for small, but difficult 3D data sets.