Learning Unorthogonalized Matrices for Rotation Estimation
Not on the map. Why?
Abstract (arXiv)
Estimating 3D rotations is a common procedure for 3D computer vision. The accuracy depends heavily on the rotation representation. One form of representation -- rotation matrices -- is popular due to its continuity, especially for pose estimation tasks. The learning process usually incorporates orthogonalization to ensure orthonormal matrices. Our work reveals, through gradient analysis, that common orthogonalization procedures based on the Gram-Schmidt process and singular value decomposition will slow down training efficiency. To this end, we advocate removing orthogonalization from the learning process and learning unorthogonalized `Pseudo' Rotation Matrices (PRoM). An optimization analysis shows that PRoM converges faster and to a better solution. By replacing the orthogonalization incorporated representation with our proposed PRoM in various rotation-related tasks, we achieve state-of-the-art results on large-scale benchmarks for human pose estimation.
Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Builds on 28 of its 39 references that are in the map. First posted on arXiv 1 Dec 2023.
Builds on
28 of 39 references.
References · 28
- 2015Deep Residual Learning for Image Recognitioncited by 52,096
- 2014Microsoft COCO: Common Objects in Contextcited by 17,608
- 2015ShapeNet: An Information-Rich 3D Model Repositorycited by 2,897
- 2018On the Continuity of Rotation Representations in Neural Networkscited by 1,214
- 2017End-to-end Recovery of Human Shape and Posecited by 885
- 2019Learning to Reconstruct 3D Human Pose and Shape via Model-fitting in the Loopcited by 627
- 2019VIBE: Video Inference for Human Body Pose and Shape Estimationcited by 620
- 2016Monocular 3D Human Pose Estimation In The Wild Using Improved CNN Supervisioncited by 494
- 2022ViTPose: Simple Vision Transformer Baselines for Human Pose Estimationcited by 456
- 2020End-to-End Human Pose and Mesh Reconstruction with Transformerscited by 418
- 2019Learning joint reconstruction of hands and manipulated objectscited by 378
- 2021PARE: Part Attention Regressor for 3D Human Body Estimationcited by 337
16 more
- 2020HybrIK: A Hybrid Analytical-Neural Inverse Kinematics Solution for 3D Human Pose and Shape Estimationcited by 307
- 2020I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Imagecited by 274
- 2019FreiHAND: A Dataset for Markerless Capture of Hand Pose and Shape from Single RGB Imagescited by 271
- 2020InterHand2.6M: A Dataset and Baseline for 3D Interacting Hand Pose Estimation from a Single RGB Imagecited by 245
- 2022CLIFF: Carrying Location Information in Full Frames into Human Pose and Shape Estimationcited by 238
- 2021AGORA: Avatars in Geography Optimized for Regression Analysiscited by 205
- 2020Beyond Static Features for Temporally Consistent 3D Human Pose and Shape from a Videocited by 163
- 2020Exemplar Fine-Tuning for 3D Human Model Fitting Towards In-the-Wild 3D Human Pose Estimationcited by 113
- 2023How Does Information Bottleneck Help Deep Learning?cited by 93
- 2020An Analysis of SVD for Deep Rotation Estimationcited by 86
- 2020On the Almost Sure Convergence of Stochastic Gradient Descent in Non-Convex Problemscited by 80
- 2019Convergence rates for the stochastic gradient descent method for non-convex objective functionscited by 66
- 2019Stochastic Gradient Descent for Nonconvex Learning without Bounded Gradient Assumptionscited by 48
- 2021Robust Differentiable SVDcited by 19
- 2018Depth with Nonlinearity Creates No Bad Local Minima in ResNetscited by 17
- 2021Understanding Dynamics of Nonlinear Representation Learning and Its Applicationcited by 11
Built on by
Citing papers per year: 2024: 1; 2025: 0; 2026: 2.
3 papers.