Research map / 3D vision / Depth and stereo
Continuous Pose for Monocular Cameras in Neural Implicit Representation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Represents monocular camera poses as an implicit neural function of time (PoseNet) optimized jointly with scene INRs, applied to NeRF from noisy poses, event-based NeRF, and NeRF-SLAM with IMU fusion and intrinsic motion frames.
Problem
Camera pose estimation in NeRF and SLAM is typically done on discrete keyframe poses, which ignores inter-frame continuity and handles asynchronous events or high-frequency IMU data poorly.
Approach
Map timestamps to SE(3) poses with an MLP (PoseNet, sinusoidal time encoding), optimize its weights jointly with the scene representation, and decompose motion into a slowly varying reference and a low-DOF intrinsic motion with an L0-relaxed DOF loss.
Key results
- On Replica, the intrinsic-motion variant reaches 0.49 cm average ATE vs 1.06 cm for NICE-SLAM.
- On TUM-RGBD, the intrinsic-motion variant reaches 2.48 cm average ATE vs 13.28 cm for NICE-SLAM.
- On EuRoC, IMU tight-coupling gives 6.16 average ATE vs 14.6 for VINS-MONO and 6.56 for NICE-SLAM on the sequences where it is reported.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace discrete pose refinement with continuous PoseNet optimization | NeRF with inaccurate camera poses · real data average | PSNR: 17.50 -> 19.58 | Helped |
| use continuous PoseNet instead of discrete BARF pose refinement | NeRF with inaccurate camera poses · real data average | Rotation error: 18.44 -> 0.446 | Helped |
| use continuous PoseNet instead of discrete BARF pose refinement | NeRF with inaccurate camera poses · real data average | Translation error: 10.777 -> 0.320 | Helped |
| add continuous PoseNet pose estimation to EventNeRF | EventNeRF · Chair and Hotdog interpolation experiments | Translation Error: 3.66 -> 1.74; 3.66 -> 2.42 | Helped |
| use continuous PoseNet tracking | NICE-SLAM · Replica tracking average | ATE-RMSE (cm): 1.06 -> 0.54 | Helped |
| add intrinsic low-DOF motion representation (ablation) | PoseNet tracking · Replica tracking average | ATE-RMSE (cm): 0.54 -> 0.49 | Helped |
| optimize relative pose in world coordinates (ablation) | PoseNet tracking · Replica tracking average | ATE-RMSE (cm): 0.54 -> 0.58 | Hurt |
| add intrinsic low-DOF motion representation (ablation) | PoseNet tracking · ScanNet tracking average | ATE-RMSE (cm): 9.86 -> 8.93 | Helped |
| add intrinsic low-DOF motion representation (ablation) | PoseNet tracking · TUM-RGBD tracking average | ATE-RMSE (cm): 4.70 -> 2.48 | Helped |
| use decoupled rotation and translation networks rather than a coupled representation (ablation) | PoseNet · NeRF pose-network architecture ablation | PSNR: 27.25 -> 27.30 | No clear effect |
| use sinusoidal embedding dimension 5 rather than dimension 2 (ablation) | PoseNet · NeRF pose-network architecture ablation | PSNR: 14.05 -> 27.25 | Helped |
| use sinusoidal embedding dimension 5 rather than dimension 10 (ablation) | PoseNet · NeRF pose-network architecture ablation | PSNR: 24.88 -> 27.25 | Helped |
Surprising findings
- Intrinsic-motion gains on ScanNet are small, which the authors attribute to noisy ground-truth poses.
Limitations
- Intrinsic-motion gain is small on ScanNet, attributed to noisy ground-truth poses.
- Random coordinate-frame tracking on Replica is unstable and fails in several sequences.
Experimental setup
Datasets: LLFF Lego chair hotdog Replica ScanNet TUM-RGBD EuRoC
Benchmarks: LLFF Replica ScanNet TUM-RGBD EuRoC
Metrics: PSNR SSIM LPIPS ATE-RMSE rotation error translation error corner error
Baselines: BARF NICE-SLAM EventNeRF Vox-Fusion ESLAM DI-Fusion VINS-MONO ORB-SLAM DROID-SLAM B-spline
Builds on: BARF NICE-SLAM EventNeRF NeRF
Closest papers by meaning
- Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video (CVPR 2025 · similarity 0.87)
- Free-Moving Object Reconstruction and Pose Estimation with Virtual Camera (AAAI 2025 · similarity 0.85)
- PLGSLAM: Progressive Neural Scene Represenation with Local to Global Bundle Adjustment (CVPR 2024 · similarity 0.85)
- Implicit Zoo: A Large-Scale Dataset of Neural Implicit Functions for 2D Images and 3D Scenes (NeurIPS 2024 · similarity 0.84)
- MegaSaM: Accurate, Fast and Robust Structure and Motion from Casual Dynamic Videos (CVPR 2025 · similarity 0.84)
- A Simple and Effective Point-based Network for Event Camera 6-DOFs Pose Relocalization (CVPR 2024 · similarity 0.84)
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.