WBF-ODAL: Weighted Boxes Fusion for 3D Object Detection from Automotive LiDAR Point Clouds

The use of Convolutional Neural Networks (CNNs) is the state-of-the-art for 3D object detection from automotive/vehicle LiDAR point clouds. However, not all models perform uniformly well across all classes. Hence, ensemble-based solutions are used where a mixture of experts (ME) approach has shown some promise. There is limited work on using the weighting function in this context, even though it is widely used in various other problem statements in computer vision. We propose the use of Weighted Boxes Fusion (WBF) for Object Detection from Automotive LiDAR point clouds (ODAL). We refer to our novel end-to-end workflow as WBF-ODAL. Our experiments on the nuScenes datasets using two different ensembles demonstrate that WBF-ODAL outperforms ME-ODAL for most classes.