Aggregated Time Series Features in a Voxel-Based Network Architecture

Using point cloud sequences is a popular way to harness the additional information represented in the time domain in order to enhance the performance of 3D object detector neural networks. However, it is not trivial to decide which abstraction level should the additional information presented to the network, or what is the point in the architecture, where aggregating the additional information is most beneficial. In this article, the authors propose various voxel-based networks and analyze their performance in relation to the abstraction level of the time series data. During the evaluation, the authors examine the object detection performance of a popular voxel-based neural network with its original architecture and several variants where the time domain related features were propagated through the network and aggregated at different stages of processing. Based on the evaluation results, a conclusion is drawn regarding the abstraction level at which the time-series aggregation step is performed in order to improve the performance of the baseline voxel-based detector.