MaskBEV: Joint Object Detection and Footprint Completion for Bird's-Eye View 3D Point Clouds

Recent works in object detection in LiDAR point clouds mostly focus on predicting bounding boxes around objects. This prediction is commonly achieved using anchor-based or anchor-free detectors that predict bounding boxes, requiring significant explicit prior knowledge about the objects to work properly. To remedy these limitations, we propose MaskBEV, a bird's-eye view (BEV) mask-based object detector neural architecture. MaskBEV predicts a set of BEV instance masks that represent the footprints of detected objects. Moreover, our approach allows object detection and footprint completion in a single pass. MaskBEV also reformulates the detection problem purely in terms of classification, doing away with regression usually done to predict bounding boxes. We evaluate the performance of MaskBEV on both SemanticKITTI and KITTI datasets while analyzing the architecture advantages and limitations.

I. Introduction

Object detection in 3D point clouds is crucial to many ap-plications of robotics and autonomous vehicles. Point clouds, captured by sensors such as LiDARs, provide accurate 3D information about the system's surroundings. However, it is more difficult to process them with deep neural networks in order to extract actionable semantic information. Indeed, point clouds, unlike images that are a dense regular grid of pixels, are irregular, unstructured and unordered [1]. Moreover, LiDAR point clouds suffer from multiple types of occlusion and signal miss [2]. One type of occlusion is external-occlusion, which is caused by obstacles blocking the laser from reaching the objects. Self-occlusion happens when an object's near side hides its far side. It is inevitable and will affect every object in a LiDAR scan. Signal miss can be caused by reflective materials reflecting the laser beam away from the sensor or by low reflectance. This often leads to objects appearing incomplete in LiDAR scans.