Object Detection using PointPillars with Modified DarkNet53 as Backbone

Three-dimensional (3D) object detection using LIDAR data is a crucial aspect in building autonomous vehicles. This study focuses on investigating the effect of modifying the backbone network for deep convolutional neural networks (DCNN) on object detection accuracy and total number of learnable parameters. The PointPillars network is employed in our experiments, which boasts a straightforward design, fast processing capabilities, and the flexibility to easily incorporate additional features. The 2D CNN network used as backbone in the benchmark PointPillars model has been replaced by a variant of Darknet53 in the proposed work. It is demonstrated that by modifying the original backbone network of the PointPillars model, there is significant improvement in the detection accuracy for higher confidence thresholds of 75% and 95% with almost half the number of learnable parameters.