Ma Xinzhu ; Zhang Yinmin ; Xu Dan ; Zhou Dongzhan ; Yi Shuai ; Li Haojie ; Ouyang Wanli

Abstract

Estimating 3D bounding boxes from monocular images is an essential componentin autonomous driving, while accurate 3D object detection from this kind ofdata is very challenging. In this work, by intensive diagnosis experiments, wequantify the impact introduced by each sub-task and found the localizationerror' is the vital factor in restricting monocular 3D detection. Besides, wealso investigate the underlying reasons behind localization errors, analyze theissues they might bring, and propose three strategies. First, we revisit themisalignment between the center of the 2D bounding box and the projected centerof the 3D object, which is a vital factor leading to low localization accuracy.Second, we observe that accurately localizing distant objects with existingtechnologies is almost impossible, while those samples will mislead the learnednetwork. To this end, we propose to remove such samples from the training setfor improving the overall performance of the detector. Lastly, we also proposea novel 3D IoU oriented loss for the size estimation of the object, which isnot affected bylocalization error'. We conduct extensive experiments on theKITTI dataset, where the proposed method achieves real-time detection andoutperforms previous methods by a large margin. The code will be made availableat: https://github.com/xinzhuma/monodle.