人工智能培训

搜索

深度学习论文:跨域3D等变图像嵌入(Cross-Domain 3D Equivariant Image Embeddings)

[复制链接]
nite_p007 发表于 4 天前 | 显示全部楼层 |阅读模式
nite_p007 4 天前 621 0 显示全部楼层
深度学习论文:跨域3D等变图像嵌入(Cross-Domain 3D Equivariant Image Embeddings)最近引入了球形卷积网络,因为工具可以容忍3D形状的强大特征表示。球形CNN与三维旋转等效,使其非常适用于可在任意方向观察到3D数据的应用。在本文中,我们学习了具有类似等变结构的2Dimage嵌入:嵌入三维对象的图像应该与对象的旋转进行通信。我们将2D图像的跨域嵌入引入球形CNN潜在空间。我们的模型仅通过从用于3D形状分类的球形CNN获得的目标嵌入来监督。训练的模型学习编码具有3D形状属性的图像,并且与观察对象的3D旋转等效。我们表明,只学习具有适当几何结构的图像的丰富嵌入本身就足以应​​对许多应用。来自两个不同应用的证据,即相对姿态估计和新颖视图合成,证明了等变量嵌入对于两个应用都是足够的,而不需要任何特定于任务的监督训练。
Spherical convolutional networks have been introduced recently as tools tolearn powerful feature representations of 3D shapes.Spherical CNNs areequivariant to 3D rotations making them ideally suited for applications where3D data may be observed in arbitrary orientations.In this paper we learn 2Dimage embeddings with a similar equivariant structure: embedding the image of a3D object should commute with rotations of the object.We introduce across-domain embedding from 2D images into a spherical CNN latent space.Ourmodel is supervised only by target embeddings obtained from a spherical CNNpretrained for 3D shape classification.The trained model learns to encodeimages with 3D shape properties and is equivariant to 3D rotations of theobserved object.We show that learning only a rich embedding for images withappropriate geometric structure is in and of itself sufficient for tacklingnumerous applications.Evidence from two different applications, relative poseestimation and novel view synthesis, demonstrates that equivariant embeddingsare sufficient for both applications without requiring any task-specificsupervised training.深度学习论文:跨域3D等变图像嵌入(Cross-Domain 3D Equivariant Image Embeddings) y6cycqnY0gqkyKkv.jpg
URL地址:https://arxiv.org/abs/1812.02716     ----pdf下载地址:https://arxiv.org/pdf/1812.02716    ----深度学习论文:跨域3D等变图像嵌入(Cross-Domain 3D Equivariant Image Embeddings)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

nite_p007当前离线
新手上路

查看:621 | 回复:0

快速回复 返回顶部 返回列表