Rigidity Preserving Image Transformations and Equivariance in Perspective
Brynte, Lucas, Bökman, Georg, Flinth, Axel, Kahl, Fredrik
–arXiv.org Artificial Intelligence
We characterize the class of image plane transformations which realize rigid camera motions and call these transformations `rigidity preserving'. In particular, 2D translations of pinhole images are not rigidity preserving. Hence, when using CNNs for 3D inference tasks, it can be beneficial to modify the inductive bias from equivariance towards translations to equivariance towards rigidity preserving transformations. We investigate how equivariance with respect to rigidity preserving transformations can be approximated in CNNs, and test our ideas on both 6D object pose estimation and visual localization. Experimentally, we improve on several competitive baselines.
arXiv.org Artificial Intelligence
Oct-13-2022
- Country:
- Europe
- Switzerland (0.04)
- United Kingdom > England
- Cambridgeshire > Cambridge (0.04)
- Sweden > Västerbotten County
- Umeå (0.04)
- Asia > Japan
- Honshū > Tōhoku > Fukushima Prefecture > Fukushima (0.04)
- Europe
- Genre:
- Research Report > New Finding (0.67)
- Industry:
- Media
- Television (0.48)
- Photography (0.48)
- Film (0.48)
- Media
- Technology: