作者：Mehmet Aygün Oisin Mac Aodha
我们介绍了SAOR，这是一种从野外拍摄的单个图像中估计关节物体的3D形状、纹理和视点的新方法。与以前依赖预定义的特定类别3D模板或定制的3D骨架的方法不同，SAOR学习使用无骨架的基于零件的模型从单个视图图像集合中表达形状，而不需要任何3D对象形状先验。为了防止不适定的解决方案，我们提出了利用解纠缠的对象形状变形和关节的跨实例一致性损失。这得益于一种新的基于轮廓的采样机制，以增强训练过程中的视点多样性。我们的方法只需要在训练过程中从现成的预训练网络中估计出物体轮廓和相对深度图。在推理时，给定一个单一的视图图像，它可以有效地输出显式网格表示。与相关研究相比，我们对具有挑战性的四足动物获得了改进的定性和定量结果
We introduce SAOR, a novel approach for estimating the 3D shape, texture, andviewpoint of an articulated object from a single image captured in the wild.Unlike prior approaches that rely on pre-defined category-specific 3D templatesor tailored 3D skeletons, SAOR learns to articulate shapes from single-viewimage collections with a skeleton-free part-based model without requiring any3D object shape priors. To prevent ill-posed solutions, we propose across-instance consistency loss that exploits disentangled object shapedeformation and articulation. This is helped by a new silhouette-based samplingmechanism to enhance viewpoint diversity during training. Our method onlyrequires estimated object silhouettes and relative depth maps fromoff-the-shelf pre-trained networks during training. At inference time, given asingle-view image, it efficiently outputs an explicit mesh representation. Weobtain improved qualitative and quantitative results on challenging quadrupedanimals compared to relevant existing work.
论文链接：http://arxiv.org/pdf/2303.13514v1
更多计算机论文：http://cspaper.cn/