Mobile Object Detection Without Labels with Camera-LiDAR Fusion for Roadside Perception

Abstract

In the vehicle-road collaboration scenario, the model performance fully rely on high-quality human-annotated data in the extensive deployment of roadside. But the cost of humanannotated labels is expensive. In this paper, we propose a novel mobile object detection method which can generate high accurate 3D object labels from unlabeled point could and images. The method mainly consists of two modules: First, we leverage combination of ephemeral points from point cloud and optical flow map from image to obtain initial labels, then we use these initial labels to train a high-precision detector via several self-training. The experimental results show that our method can effectively train a high accurate mobile object detector without relying on any manual labeling.