Efficient combination of classifiers for 3D action recognition

Abstract

The popular task of 3D human action recognition is almost exclusively solved by training deep-learning classifiers. To achieve high recognition accuracy, input 3D actions are often pre-processed by various normalization or augmentation techniques. However, it is not computationally feasible to train a classifier for each possible variant of training data to select the best-performing combination of pre-processing techniques for a given dataset. In this paper, we propose an evaluation procedure that determines the best combination in a very efficient way. In particular, we only train one independent classifier for each available pre-processing technique and estimate the accuracy of a specific combination by efficient fusion of the corresponding classification results based on a strict majority vote rule. In addition, for the best-ranked combination, we can retrospectively apply the normalized/augmented variants of input data to train only a single classifier. This enables to decide whether it is generally better to train a single model, or rather a set of independent classifiers whose results are fused within the classification phase. We evaluate the experiments on single-subject as well as person-interaction datasets of 3D skeleton sequences and all combinations of up to 16 normalization and augmentation techniques, some of which are proposed in this paper.