Precision Vehicle Pose Estimation with Uncertainty-aware Neural Network

This study presents a neural network designed for precise vehicle pose estimation from single images in complex settings. Utilising neural network backbones known for accurate human pose keypoint detection, our architecture effectively localises vehicle characteristic points. Task-specific modules estimate point coordinates and quality for pose computation. Training on ApolloCar3D with auto-generated 3D labels, our approach achieves the high pose estimation accuracy. We highlight the crucial role of accurate keypoint detection in addressing single-view geometry ambiguities, enhancing pose estimation precision.