Research map / Optimization / Nonconvex and smooth optimization
A Game of Bundle Adjustment - Learning Efficient Convergence
Amir Belder, Refael Vivanti, Ayellet Tal
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 |
|---|---|---|---|
| learn lambda holistically with RL (SAC) instead of non-holistic Zero-Net supervised scheme (ablation) | Ours + classic · BAL | iterations to converge: 8 -> 4 | Improved |
| use state size 5 instead of 1 (consecutive estimation errors) (ablation) | Ours + classic · BAL | iterations to converge: 7.5 -> 4 | Improved |
| increase state size from 5 to 20 (ablation) | Ours + classic · BAL | iterations to converge: 4 -> 5 | Hurt |
| use estimation error as state and time as reward instead of reversed roles (ablation) | Ours + classic · BAL | iterations / success rate: 10 -> 4 iterations; 70% -> 100% success | Improved |
| use iteration-duration reward instead of constant negative reward (ablation) | Ours + classic · BAL | iterations to converge: 5 -> 4 | Improved |
| use iteration-duration reward instead of reward reduction (ablation) | Ours + classic · BAL | iterations to converge: 4.2 -> 4 | Improved |
| use learned dynamic lambda instead of a fixed lambda schedule (ablation) | Ours + classic · BAL | iterations to converge: 12 -> 4 | Improved |
| train the agent on small synthetic random-point scenes instead of KITTI/BAL (ablation) | Ours + classic · KITTI and BAL | iterations to converge: 14 -> 17 (KITTI); 4 -> 5 (BAL) | No clear effect |
| apply learned damping on very small BA problems (5 points, 5 poses) | Ours + classic · synthetic small problems | iterations / time: 5 -> 4 iterations but about 0.5 s slower overall | Hurt |
| use the learned damping agent with a high tolerance (0.1) | Ours + classic · BAL performance profile | solve time | No clear effect |
Closest papers by meaning
- Learning To Bundle-Adjust: A Graph Network Approach to Faster Optimization of Bundle Adjustment for Vehicular SLAM (ICCV 2021 · similarity 0.86)
- From Variance to Veracity: Unbundling and Mitigating Gradient Variance in Differentiable Bundle Adjustment Layers (CVPR 2024 · similarity 0.84)
- Stochastic Bundle Adjustment for Efficient and Scalable 3D Reconstruction (ECCV 2020 · similarity 0.83)
- ICE-BA: Incremental, Consistent and Efficient Bundle Adjustment for Visual-Inertial SLAM (CVPR 2018 · similarity 0.82)
- Active Neural Localization (ICLR 2018 · similarity 0.82)
- Learning Dynamics under Environmental Constraints via Measurement-Induced Bundle Structures (ICML 2025 · similarity 0.82)
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.
