Yang Zhen ; Dong Yanpeng ; Wang Heng ; Ma Lichao ; Cui Zijian ; Liu Qi ; Pei Haoran

Abstract

Multi-sensor fusion significantly enhances the accuracy and robustness of 3Dsemantic occupancy prediction, which is crucial for autonomous driving androbotics. However, most existing approaches depend on large image resolutionsand complex networks to achieve top performance, hindering their application inpractical scenarios. Additionally, most multi-sensor fusion approaches focus onimproving fusion features while overlooking the exploration of supervisionstrategies for these features. To this end, we propose DAOcc, a novelmulti-modal occupancy prediction framework that leverages 3D object detectionsupervision to assist in achieving superior performance, while using adeployment-friendly image feature extraction network and practical input imageresolution. Furthermore, we introduce a BEV View Range Extension strategy tomitigate the adverse effects of reduced image resolution. Experimental resultsshow that DAOcc achieves new state-of-the-art performance on the Occ3D-nuScenesand SurroundOcc benchmarks, and surpasses other methods by a significant marginwhile using only ResNet50 and 256*704 input image resolution. Code will be madeavailable at https://github.com/AlphaPlusTT/DAOcc.