Research map / 3D vision / Human pose estimation
Deep Feedback Inverse Problem Solver
Wei-Chiu Ma, Shenlong Wang, Jiayuan Gu, Sivabalan Manivasagam, Antonio Torralba, Raquel Urtasun
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the forward-process feedback signal (compare rendered and observed images) to the regressor | deep regression network · 6-DoF pose estimation (ShapeNet silhouettes) | mean rotation error (deg): 10.07 -> 2.64 | Improved |
| add the forward-process feedback signal | deep regression network · 6-DoF pose estimation (ShapeNet silhouettes) | outlier rate (%): 5 -> 2.6 | Improved |
| add the feedback signal | deep regression (Janner et al.) · Illumination estimation (directional light) | mean error (MSE): 0.067 -> 0.052 | Improved |
| add the feedback signal | deep regression (Janner et al.) · Illumination estimation (point light) | mean error (MSE): 0.111 -> 0.084 | Improved |
| replace differentiable-renderer optimization with the learned feedback update | SoftRas optimization · 6-DoF pose estimation (ShapeNet silhouettes) | outlier rate (%): 8.03 -> 2.6 | Improved |
| replace differentiable-renderer optimization with the learned feedback update | SoftRas optimization · 6-DoF pose estimation (ShapeNet silhouettes) | mean translation error: 0.05 -> 0.02 | Improved |
| replace Adam optimization with the learned feedback update | Adam on joint position error · Inverse kinematics (CMU MoCap) | mean position error (cm): 0.04 -> 0.64 | Hurt |
| replace Adam optimization with the learned feedback update | Adam on joint position error · Inverse kinematics (CMU MoCap) | mean rotation error (deg): 7.96 -> 0.88 | Improved |
| use rotations farther from the training range at test time | Deep Feedback (Ours) · 6-DoF pose, unseen rotations | mean rotation error (deg): 11.33 -> 24.14 | Hurt |
| use the feedback network as initialization for SoftRas | SoftRas · 6-DoF pose estimation (ShapeNet silhouettes) | error: reduce the error by more than 50% | Improved |
| replace shared update weights with per-step (stage-wise) networks (ablation) | Deep Feedback (Ours) · 6-DoF pose estimation | error | Improved |
| add deep supervision on every intermediate step (ablation) | Deep Feedback (Ours) · training | convergence speed | Improved |
Closest papers by meaning
- Benchmarking Deep Inverse Models over time, and the Neural-Adjoint method (NeurIPS 2020 · similarity 0.88)
- An analytic theory of convolutional neural network inverse problems solvers (ICML 2026 · similarity 0.88)
- One Network to Solve Them All -- Solving Linear Inverse Problems Using Deep Projection Models (ICCV 2017 · similarity 0.88)
- Solving Inverse Problems via Diffusion Optimal Control (NeurIPS 2024 · similarity 0.87)
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.86)
- NPN: Non-Linear Projections of the Null-Space for Imaging Inverse Problems (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
