Ab­stract

Con­vo­lu­tional Neural Net­works (CNNs) achieve high ac­cu­racy in vi­sion-based ob­ject de­tec­tion tasks. For their usage in the au­to­mo­tive do­main, CNNs have to be ro­bust against var­i­ous kinds of nat­ural dis­tor­tions caused by dif­fer­ent weather con­di­tions while state-of-the-art datasets like KITTI lack these chal­leng­ing sce­nar­ios. Our ap­proach au­to­mat­i­cally iden­ti­fies cor­ner cases where CNNs fail and im­proves their ro­bust­ness by au­to­mated aug­men­ta­tion of the train­ing data with syn­thetic rain vari­a­tions in­clud­ing falling rain with bright­ness re­duc­tion as well as rain­drops on the wind­shield. Our method achieves higher per­for­mance upon val­i­da­tion against a real rain dataset com­pared with state-of-the art data aug­men­ta­tion tech­niques like Gauss­ian noise (GN) or Salt-and-Pep­per noise (SPN).