Sparse PointPillars: Exploiting Sparsity in Birds-Eye-View Object Detection

Bird's Eye View (BEV) is a popular representation for processing 3D point clouds, and by its nature is fundamentally sparse. Motivated by the computational limitations of mobile robot platforms, we take a fast high-performance BEV 3D object detector - PointPillars - and modify its backbone to exploit this sparsity, leading to decreased runtimes. We present preliminary results demonstrating decreased runtimes with either the same performance or a modest decrease in performance, which we anticipate will be remedied by model specific hyperparameter tuning. Our work is a first step towards a new class of 3D object detectors that exploit sparsity throughout their entire pipeline in order to reduce runtime and resource usage while maintaining good detection performance.