Lu Yan ; Ma Xinzhu ; Yang Lei ; Zhang Tianzhu ; Liu Yating ; Chu Qi ; Yan Junjie ; Ouyang Wanli

Abstract

Geometry Projection is a powerful depth estimation method in monocular 3Dobject detection. It estimates depth dependent on heights, which introducesmathematical priors into the deep model. But projection process also introducesthe error amplification problem, in which the error of the estimated heightwill be amplified and reflected greatly at the output depth. This propertyleads to uncontrollable depth inferences and also damages the trainingefficiency. In this paper, we propose a Geometry Uncertainty Projection Network(GUP Net) to tackle the error amplification problem at both inference andtraining stages. Specifically, a GUP module is proposed to obtains thegeometry-guided uncertainty of the inferred depth, which not only provides highreliable confidence for each depth but also benefits depth learning.Furthermore, at the training stage, we propose a Hierarchical Task Learningstrategy to reduce the instability caused by error amplification. This learningalgorithm monitors the learning situation of each task by a proposed indicatorand adaptively assigns the proper loss weights for different tasks according totheir pre-tasks situation. Based on that, each task starts learning only whenits pre-tasks are learned well, which can significantly improve the stabilityand efficiency of the training process. Extensive experiments demonstrate theeffectiveness of the proposed method. The overall model can infer more reliableobject depth than existing methods and outperforms the state-of-the-artimage-based monocular 3D detectors by 3.74% and 4.7% AP40 of the car andpedestrian categories on the KITTI benchmark.