3DTDesc: learning local features using 2D and 3D cues

Abstract

Pairwise frame registration with sparse geometric local features on real-world depth images is not particularly robust due to the low resolution and incomplete nature of the 3D scan data. Moreover, there might be many regions with similar geometric information. In this paper, we present 3DTDesc, a data-driven descriptor which closely combines both 2D texture and 3D geometric information for frame registration. The proposed descriptor is learned directly from color point clouds, which is time-efficient and provides robust and accurate geometric feature matching in a variety of settings. The texture information and the geometric information closely interact in the fusing network, which are complements of each other in situations of textureless regions or regions with similar geometric information and different texture information. We also propose a multi-scale 3DTDesc to further improve the performance of the feature matching. The effectiveness and efficiency of our proposed 3DTDesc are demonstrated by extensive experimental results on challenging RGB-D datasets and various ablation studies.