Research map / 3D vision / NeRF and radiance fields

GoLF-NRT: Integrating Global Context and Local Geometry for Few-Shot View Synthesis

CVPR 2025 · 3D vision › NeRF and radiance fields · approach: Few-shot neural rendering architectures · 2 citations (OpenAlex)

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 GoLF-NRT, a generalizable NeRF that fuses a sparse-attention 3D transformer global context with epipolar local geometry features plus kernel-regression adaptive sampling, for few-shot novel view synthesis from 1-3 views.

Problem

Generalizable NeRF methods degrade sharply when given very few input views (1-3), producing over-smoothed textures and artifacts in occluded or reflective regions.

Approach

Combine a coarse global scene context from a cross-view 3D transformer with multi-axis sparse attention, which then guides the selection of epipolar local geometry features; attention weights are converted to a smooth PDF via kernel regression for adaptive sampling.

Key results

• With 3 input views on LLFF, GoLF-NRT reaches 24.20 PSNR, 0.75 dB above CaesarNeRF (23.45).
• With 1 input view on Shiny, GoLF-NRT improves over CaesarNeRF by 1.39 dB PSNR.
• Ablation on LLFF (3 views): global context adds 0.27 dB, and kernel-regression adaptive sampling adds 0.32 dB over attention-only sampling, for a total gain of 0.92 dB over local-only (23.28 to 24.20).

Measured effects reported in the paper

Limitations

• Performance still degrades with fewer input views, especially at 1 view on LLFF and Blender.
• Training data sources are constrained to a specific mixture of synthetic and real scenes; generalization beyond evaluated datasets is not extensively studied.

Experimental setup

Datasets: Google Scanned Objects RealEstate10K Spaces LLFF-style real handheld scenes

Benchmarks: LLFF Blender Shiny

Metrics: PSNR SSIM LPIPS

Baselines: PixelNeRF IBRNet MVSNeRF MatchNeRF NeuRay GPNR GNT CaesarNeRF EVE-NeRF pixelSplat MVSGaussian

Builds on: NeRF IBRNet GNT CaesarNeRF EVE-NeRF

Closest papers by meaning

• Is Attention All That NeRF Needs? (ICLR 2023 · similarity 0.90)
• FrugalNeRF: Fast Convergence for Extreme Few-shot Novel View Synthesis without Learned Priors (CVPR 2025 · similarity 0.90)
• NeRF-VPT: Learning Novel View Representations with Neural Radiance Fields via View Prompt Tuning (AAAI 2024 · similarity 0.90)
• Few-Shot Neural Radiance Fields under Unconstrained Illumination (AAAI 2024 · similarity 0.89)
• CVT-xRF: Contrastive In-Voxel Transformer for 3D Consistent Radiance Fields from Sparse Inputs (CVPR 2024 · similarity 0.88)
• Learning Robust Generalizable Radiance Field with Visibility and Feature Augmented Point Representation (ICLR 2024 · similarity 0.88)