Scholarlink

<p>Sparse-view neural radiance fields (NeRFs) employed for reconstructing large-scale ancient architecture face several challenges, including geometric misalignment, inadequate synergy between high and low frequencies, and material detail distortion. To address these limitations, we propose a coordinate feature gate-based collaborative optimization and neural radiance field (CFG-NeRF) method with three novel technical contributions: (1) a multimodal injection MLP (MI-MLP) that incorporates a coordinate-based layerwise injection mechanism integrated with gradient redistribution, which dynamically balances high- and low-frequency features and replaces the single-weighted fusion process to eliminate geometric misalignment issue; (2) a feature pyramid (FP) module that enables the implementation of dynamic feature selection, bilinear resolution alignment, and adaptive weighted summation to precisely coordinate macro- and microscale structural features; (3) a cross-modal gated attention (CGA) module that generates auxiliary samples and integrates spatial-, channel-, and pixel-level attention mechanisms to calibrate reflective material features via physical constraints. Extensive experiments conducted on the Zhantan Temple dataset demonstrate state-of-the-art performance, with CFG-NeRF achieving 22.73 dB PSNR, 0.735 SSIM, and 0.174 LPIPS–surpassing the best baseline by 16.27%, 12.21%, and 33.84%, respectively. The code is available at <a href="https://github.com/wangsaiweiwsw/CFG-NeRF">https://github.com/wangsaiweiwsw/CFG-NeRF</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CFG-NeRF: coordinate-feature-gate collaborative optimization for sparse-view ancient architecture reconstruction

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Lihua Hu,
  • Saiwei Wang,
  • Jianhua Hu,
  • Xiaoling Yao,
  • Sulan Zhang

摘要

Sparse-view neural radiance fields (NeRFs) employed for reconstructing large-scale ancient architecture face several challenges, including geometric misalignment, inadequate synergy between high and low frequencies, and material detail distortion. To address these limitations, we propose a coordinate feature gate-based collaborative optimization and neural radiance field (CFG-NeRF) method with three novel technical contributions: (1) a multimodal injection MLP (MI-MLP) that incorporates a coordinate-based layerwise injection mechanism integrated with gradient redistribution, which dynamically balances high- and low-frequency features and replaces the single-weighted fusion process to eliminate geometric misalignment issue; (2) a feature pyramid (FP) module that enables the implementation of dynamic feature selection, bilinear resolution alignment, and adaptive weighted summation to precisely coordinate macro- and microscale structural features; (3) a cross-modal gated attention (CGA) module that generates auxiliary samples and integrates spatial-, channel-, and pixel-level attention mechanisms to calibrate reflective material features via physical constraints. Extensive experiments conducted on the Zhantan Temple dataset demonstrate state-of-the-art performance, with CFG-NeRF achieving 22.73 dB PSNR, 0.735 SSIM, and 0.174 LPIPS–surpassing the best baseline by 16.27%, 12.21%, and 33.84%, respectively. The code is available at https://github.com/wangsaiweiwsw/CFG-NeRF.