Research map / 3D vision / Human pose estimation
DeepSFM: Structure From Motion Via Deep Bundle Adjustment
Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, Xiangyang Xue
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 geometric (depth) consistency channels to cost volumes (ablation) | DeepSFM · MVS | abs_rel: 0.0813 -> 0.0698 | Improved |
| add geometric consistency channels (ablation) | DeepSFM · MVS | sq_rel: 0.0971 -> 0.0523 | Improved |
| use nearest neighbor instead of bilinear sampling for depth warping (ablation) | DeepSFM · DeMoN MVS | L1-inv: 0.023 -> 0.021 | Improved |
| use nearest neighbor instead of bilinear sampling for depth warping (ablation) | DeepSFM · DeMoN MVS | Trans error: 9.910 -> 9.881 | Improved |
| replace direct visual odometry pose update with P-CV (ablation) | DeepSFM · DeMoN iterations | depth and pose error | Improved |
| iterate refinement from DeMoN initialization up to iteration 4 (ablation) | DeepSFM · DeMoN iterations (test) | abs relative: 0.254 -> 0.126 | Improved |
| run more than 4 test iterations (up to 20) (ablation) | DeepSFM · DeMoN iterations (test) | abs relative: 0.126 -> 0.120 | No clear effect |
| add scale invariant gradient loss (ablation) | DeepSFM · MVS | abs_rel: 0.0698 -> 0.0712 | No clear effect |
| add small Gaussian noise to initialization during fine-tuning (ablation) | DeepSFM · MVS | depth and pose accuracy | Hurt |
| use translation bin size 0.10 instead of 0.05 (ablation) | DeepSFM · MVS | Rot error: 3.308 -> 2.825 | Improved |
| use rotation bin 0.03 instead of 0.07 (ablation) | DeepSFM · MVS | Rot error: 3.024 -> 2.825 | Improved |
| use DeepSFM instead of COLMAP | COLMAP · ETH3D | sq_rel: 36.71 -> 0.278 | Improved |
Later papers that compared with this method
2 later accepted papers report DeepSfM as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Multi-body Depth and Camera Pose Estimation from Multiple Views (ICCV 2023 · better on 3 of 3 shared results)
- Deep Two-View Structure-From-Motion Revisited (CVPR 2021 · better on 6 of 6 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- DeMoN — DeMoN: Depth and Motion Network for Learning Monocular Stereo (CVPR 2017) · better on 3 of 3 shared results
- BA-Net — BA-Net: Dense Bundle Adjustment Networks (ICLR 2019) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DeepSFM: Robust Deep Iterative Refinement for Structure From Motion (TPAMI 2023 · similarity 0.96)
- VGGSfM: Visual Geometry Grounded Deep Structure From Motion (CVPR 2024 · similarity 0.90)
- Learning Structure-from-Motion with Graph Attention Networks (CVPR 2024 · similarity 0.89)
- Deep Non-Rigid Structure-from-Motion Revisited: Canonicalization and Sequence Modeling (AAAI 2025 · similarity 0.88)
- MP-SfM: Monocular Surface Priors for Robust Structure-from-Motion (CVPR 2025 · similarity 0.86)
- Light3R-SfM: Towards Feed-forward Structure-from-Motion (CVPR 2025 · similarity 0.86)
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.
