Abstract
Monocular 3D object detection is an active research topic due to its challenging nature and broad applying prospects. In this work, a single-stage keypoint-based network, named as FADNet, is presented to address the task of monocular 3D object detection with autonomous driving as the target application. In contrast to previous keypoint-based methods which adopt identical layouts for output branches, we propose to divide the output modalities into different groups according to the estimating difficulty, whereby different groups are treated differently. To this end, a convolutional gated recurrent unit (convGRU) is embedded into our network to enable sequential feature association across the convolutional features in different groups. The purpose of sequential feature association is to improve the accuracy of harder estimations under the guidance of easier ones. It is also observed that such design contributes to the geometric consistency between 2D and 3D estimations. Another contribution of this work is the strategy of depth hint augmentation. To provide characterized depth patterns as hints for depth estimation, a dedicated depth hint module is designed to generate row-wise features named as depth hints, which are explicitly supervised in a bin-wise manner. In the training stage, the regression outputs are uniformly encoded to enable loss disentanglement. The 2D loss term is further adapted to be depth-aware for improving the detection accuracy of small objects. The contributions of this work are validated by conducting experiments and ablation study on the KITTI3D benchmark. Without utilizing depth priors, post optimization, or other refinement modules, our network achieves the performance on par with state-of-the-art methods while maintaining a decent running speed. The code is available at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2011.14589
