Part Detection for 3D Shapes via Multi-view Rendering

Learning semantic abstraction of shape via 3D region of interest

“…We compare our approach with Song et al's method [SSSW17] in our benchmark dataset. This [SSSW17] is a multi-view based method, which has good effect for the 3D surface model. However, the part beneath the surface of the 3D object cannot be abstracted, as it is not visible.…”

“…We compare our approach with Song et al's method [SSSW17] in our benchmark dataset. This [SSSW17] is a multi-view based method, which has good effect for the 3D surface model.…”

“…However, the part beneath the surface of the 3D object cannot be abstracted, as it is not visible. As shown in Figure 9, the seats and steering wheel in the car are completely occluded by the car shell and thus cannot be abstracted using this method [SSSW17]. Moreover, this method [SSSW17] only estimates the Axis Align Bounding Box (AABB) of the 3D shape's semantic parts, while our method enables more compact semantic abstraction.…”

“…We then determine the semantic category of the current face f by voting, using semantic information on K nearest neighbors. Figure 12 shows a comparison between our method and Song et al's method [SSSW17] for instance-level semantic segmentation.…”

“…Based on the contour expression, the original contour lines were fitted into three types of surface and At the same time, with the rapid development of neural networks, there are also many shape abstraction methods based on neural networks [SSSW17, TSG * 17]. Song et al [SSSW17] proposed a multi-view based shape abstraction method. Firstly,they render a 3D shape into images under different perspectives, and detect the potential parts (as boxes) in the images using Faster RCNN [RHGS15].…”