Research map / 3D vision / NeRF and radiance fields
Improving Robustness for Joint Optimization of Camera Pose and Decomposed Low-Rank Tensorial Radiance Fields
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.
Proposes a robust joint optimization of camera poses and decomposed low-rank (TensoRF-style) radiance fields using only 2D images, via Gaussian-filtered supervision and radiance fields, randomly scaled kernels, and edge-guided loss.
Problem
Joint refinement of camera poses and voxel/tensor-based NeRF geometry from 2D images is prone to sub-optimal local minima because high-frequency signals create flipping, noisy gradients for pose alignment.
Approach
Analyze 1D signal alignment via Fourier transfer function showing gradient sign flips at high frequencies; apply Gaussian filtering to both 2D supervision and the decomposed tensor radiance field (with separable component-wise convolution), using a decreasing kernel schedule, randomly scaled kernels, and edge-guided rendering loss.
Key results
- Achieves pose registration and view synthesis improvements over BARF, GARF, and HASH on NeRF-Synthetic and LLFF as reported in the paper (exact numbers not given in this excerpt).
Surprising findings
- Blurring only the 2D training images is not the best choice; restricting the spectrum of both 2D images and the 3D radiance field gives the best reconstruction quality.
- A larger Gaussian kernel does not always yield more robust pose registration because overly aggressive filtering weakens and destabilizes gradients.
Limitations
- Theoretical analysis relies on assumptions such as rapid convergence of the radiance field to temporary optima before pose updates.
- Results for GARF and HASH are taken from their papers because implementations are unavailable.
Experimental setup
Datasets: NeRF-Synthetic LLFF ImageNet
Benchmarks: NeRF-Synthetic LLFF
Metrics: rotation error translation error PSNR SSIM LPIPS warp error
Baselines: BARF GARF HASH
Builds on: BARF TensoRF NeRF
Closest papers by meaning
- Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video (CVPR 2025 · similarity 0.87)
- ICON: Incremental CONfidence for Joint Pose and Radiance Field Optimization (CVPR 2024 · similarity 0.87)
- Mask-Based Modeling for Neural Radiance Fields (ICLR 2024 · similarity 0.86)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.86)
- Synergistic Integration of Coordinate Network and Tensorial Feature for Improving Neural Radiance Fields from Sparse Inputs (ICML 2024 · similarity 0.86)
- AltNeRF: Learning Robust Neural Radiance Field via Alternating Depth-Pose Optimization (AAAI 2024 · 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.