3D-SSD: Learning hierarchical features from RGB-D images for amodal 3D object detection

摘要

This paper aims at fast and high-accuracy amodal 3D object detections inRGB-D images, which requires a compact 3D bounding box around the whole objecteven under partial observations. To avoid the time-consuming proposalspreextraction, we propose a single end-to-end framework based on the deepneural networks which hierarchically incorporates appearance and geometricfeatures from 2.5D representation to 3D objects. The depth information hashelped on reducing the output space of 3D bounding boxes into a manageable setof 3D anchor boxes with different sizes on multiple feature layers. Atprediction time, in a convolutional fashion, the network predicts scores forcategories and adjustments for locations, sizes and orientations of each 3Danchor box, which has considered multi-scale 2D features. Experiments on thechallenging SUN RGB-D datasets show that our algorithm outperforms thestate-of-the-art by 10.2 in mAP and is 88x faster than the Deep Sliding Shape.In addition, experiments suggest our algorithm even with a smaller input imagesize performs comparably but is 454x faster than the state-of-art on NYUV2datasets.