Alert button

View Generalization for Single Image Textured 3D Models

Add code
Alert button
Jun 10, 2021
Figure 1 for View Generalization for Single Image Textured 3D Models
Figure 2 for View Generalization for Single Image Textured 3D Models
Figure 3 for View Generalization for Single Image Textured 3D Models
Figure 4 for View Generalization for Single Image Textured 3D Models

Share this with someone who'll enjoy it:

Humans can easily infer the underlying 3D geometry and texture of an object only from a single 2D image. Current computer vision methods can do this, too, but suffer from view generalization problems - the models inferred tend to make poor predictions of appearance in novel views. As for generalization problems in machine learning, the difficulty is balancing single-view accuracy (cf. training error; bias) with novel view accuracy (cf. test error; variance). We describe a class of models whose geometric rigidity is easily controlled to manage this tradeoff. We describe a cycle consistency loss that improves view generalization (roughly, a model from a generated view should predict the original view well). View generalization of textures requires that models share texture information, so a car seen from the back still has headlights because other cars have headlights. We describe a cycle consistency loss that encourages model textures to be aligned, so as to encourage sharing. We compare our method against the state-of-the-art method and show both qualitative and quantitative improvements.

* CVPR 2021. Project website: https://nv-adlr.github.io/view-generalization  

Share this with someone who'll enjoy it: