Research map / 3D vision / Depth and stereo
Towards Better Generalization: Joint Depth-Pose Learning Without PoseNet
Wang Zhao, Shaohui Liu, Yezhi Shu, Yong-Jin Liu
Check 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 |
|---|---|---|---|
| Use optical-flow correspondences and two-view geometry instead of a PoseNet pose regressor, with explicit depth-scale alignment by triangulation | Joint depth-pose learning · NYUv2 depth test | rel: 0.283 -> 0.201 | Improved |
| Use optical-flow pose recovery and triangulation-based depth-scale alignment | Joint depth-pose learning · KITTI Odometry Seq. 09, stride 3 | translation error: 41.24 -> 7.21 | Improved |
| Jointly train depth-pose learning with the optical-flow module rather than training FlowNet-only | Ours optical-flow estimator · KITTI 2015 training set | All EPE: 8.97 -> 5.72 | Improved |
| Add reprojection loss Lp (ablation) | Ours depth model · KITTI Eigen test split | AbsRel: 0.135 -> 0.130 | Improved |
| Add ImageNet pretraining (ablation) | Ours depth model · KITTI Eigen test split | AbsRel: 0.130 -> 0.113 | Improved |
| Remove triangulation supervision loss Ld (ablation) | Ours depth model · KITTI Eigen test split | training convergence | Hurt |
| Replace optical flow and two-view triangulation with a PoseNet-like pose branch (ablation) | Joint depth-pose learning · NYUv2 depth test | rel: 0.201 -> 0.283 | Hurt |
| Predict pose with PoseNet using optical flow as input (ablation) | Joint depth-pose learning · NYUv2 depth test | rel: 0.201 -> 0.221 | Hurt |
| Increase NYUv2 evaluation resolution to 448×576 (ablation) | Ours depth model · NYUv2 depth test | rel: 0.201 -> 0.189 | Improved |
| Use the proposed method instead of PoseNet | TUM-RGBD visual odometry · TUM-RGBD fr3/cabinet | absolute translational RMSE (meter): 1.45 -> 1.09 | Improved |
| Use the proposed method instead of ORB-SLAM2 | TUM-RGBD visual odometry · TUM-RGBD fr2/desk | absolute translational RMSE (meter): 0.006 -> 0.52 | Hurt |
| Train on TUM-RGBD and transfer to NYUv2 rather than use the PoseNet baseline | Depth model transfer · NYUv2 | AbsRel: 0.324 -> 0.276 | Improved |
Closest papers by meaning
- Generalizing to the Open World: Deep Visual Odometry With Online Adaptation (CVPR 2021 · similarity 0.89)
- DualRefine: Self-Supervised Depth and Pose Estimation Through Iterative Epipolar Sampling and Refinement Toward Equilibrium (CVPR 2023 · similarity 0.89)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.89)
- Unsupervised Learning of Monocular Depth Estimation and Visual Odometry With Deep Feature Reconstruction (CVPR 2018 · similarity 0.89)
- Sequential Adversarial Learning for Self-Supervised Deep Visual Odometry (ICCV 2019 · similarity 0.89)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.87)
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.
