Rigidity Preserving Image Transformations and Equivariance in Perspective
Not on the map. Why?
Abstract (arXiv)
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.
Key facts. Cited by 3 papers in the map (data through 7 Oct 2026). Builds on 36 of its 62 references that are in the map. First posted on arXiv 31 Jan 2022.
Builds on
36 of 62 references.
References · 36
- 2014Microsoft COCO: Common Objects in Contextcited by 17,498
- 2014Very Deep Convolutional Networks for Large-Scale Image Recognitioncited by 15,267
- 2019RandAugment: Practical automated data augmentation with a reduced search spacecited by 2,261
- 2016Group Equivariant Convolutional Networkscited by 1,154
- 2019EfficientDet: Scalable and Efficient Object Detectioncited by 1,018
- 2018AutoAugment: Learning Augmentation Policies from Datacited by 728
- 2021E(3)-Equivariant Graph Neural Networks for Data-Efficient and Accurate Interatomic Potentialscited by 701
- 2018Spherical CNNscited by 452
- 2019Making Convolutional Networks Shift-Invariant Againcited by 390
- 2018PVNet: Pixel-wise Voting Network for 6DoF Pose Estimationcited by 385
- 2016DSAC - Differentiable RANSAC for Camera Localizationcited by 264
- 2019Gauge Equivariant Convolutional Networks and the Icosahedral CNNcited by 262
24 more
- 2017Learning SO(3) Equivariant Representations with Spherical CNNscited by 244
- 2021Back to the Feature: Learning Robust Camera Localization from Pixels to Posecited by 199
- 2018Why do deep convolutional networks generalize so poorly to small image transformations?cited by 198
- 2019DPOD: 6D Pose Object Detector and Refinercited by 189
- 2020Visual Camera Re-Localization from RGB and RGB-D Images Using DSACcited by 176
- 2020How Much Position Information Do Convolutional Neural Networks Encode?cited by 171
- 2020On Translation Invariance in CNNs: Convolutional Layers can Exploit Absolute Spatial Locationcited by 141
- 2018Universal approximations of invariant maps by neural networkscited by 127
- 2019Adversarial AutoAugmentcited by 111
- 2020HybridPose: 6D Object Pose Estimation under Hybrid Representationscited by 107
- 2019Hierarchical Scene Coordinate Classification and Regression for Visual Localizationcited by 93
- 2019Single-Stage 6D Object Pose Estimationcited by 88
- 2019Tangent Images for Mitigating Spherical Distortioncited by 79
- 2019Spherical CNNs on Unstructured Gridscited by 76
- 2019End-to-End Learnable Geometric Vision by Backpropagating PnP Optimizationcited by 65
- 2020DeepSphere: a graph-based spherical CNNcited by 65
- 2021Ground-aware Monocular 3D Object Detection for Autonomous Drivingcited by 65
- 2018Kernel Transformer Networks for Compact Spherical Convolutioncited by 56
- 2020Spin-Weighted Spherical CNNscited by 50
- 2020A Wigner-Eckart Theorem for Group Equivariant Convolution Kernelscited by 46
- 2020EfficientPose: An efficient, accurate and scalable end-to-end 6D multi object pose estimation approachcited by 46
- 2018Omnidirectional DSO: Direct Sparse Odometry with Fisheye Camerascited by 17
- 2019Convolutions on Spherical Imagescited by 12
- 2020Data augmentation with Mobius transformationscited by 9
Built on by
Citing papers per year: 2023: 2; 2024: 0; 2025: 1.
3 papers.