ME-ODAL: Mixture-of-Experts Ensemble of CNN Models for 3D Object Detection from Automotive LiDAR Point Clouds

3D object detection from automotive/vehicle LiDAR point clouds is used for environment perception which is much needed for autonomous driving. While this task is achieved using different deep learning models, not all models work well for all classes and their instances. There is limited work in the use of ensemble methods for LiDAR point cloud analysis that finds the optimal model from a set of existing models. We propose a workflow for an ensemble method for Object Detection from Automotive LiDAR point clouds (ODAL), built using convolutional neural networks (CNNs). Our proposed method is a Mixture-of-Experts method (ME-ODAL) for Object Detection from Automotive LiDAR. For ME-ODAL, we propose a regression model learned on shape descriptor features as the meta-learning model, i.e., the gating function. We tested the workflow on the nuScenes dataset using two ensembles, i.e., different sets of CNNs to compare their performance. Our experimental results show that the ensemble of models demonstrates the expected results of overall improved accuracy.