Scholarlink

<p>Achieving accurate and generalizable 3D reconstruction across diverse scenes remains a major challenge for neural radiance field methods, primarily due to the lack of explicit geometric constraints in existing generalizable frameworks. While current NeRF frameworks achieve impressive photo-realistic renderings, they struggle with scene diversity and geometric fidelity. This paper introduces DeepFusion NeRF (DFN), a Transformer-based framework that integrates monocular depth information to enhance geometric consistency and texture fidelity, specifically in cross-scene novel view synthesis. DFN employs a dual-transformer architecture, comprising a View Transformer and a Ray Transformer, to aggregate multi-view features and decode them along sampled ray points. Monocular depth estimates, provided by an auxiliary network, are fused into the feature representation to enhance geometric consistency without explicit 3D supervision. Our method includes a convolutional gated squeeze–excitation (GSE) module for improved local feature encoding and a composite multi-scale rotary positional encoding scheme for robustness to scale variations. The proposed framework fills the gap left by existing methods that rely on photometric consistency alone, thus improving robustness, especially in occluded or texture-ambiguous regions. Experiments demonstrate DFN’s superior geometric fidelity and competitive rendering quality on both single-scene and cross-scene benchmarks. Codes and models are available at <a href="https://github.com/alaneze123/DeepFusion-NeRF-DFN-.">https://github.com/alaneze123/DeepFusion-NeRF-DFN-.</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

Enhancing neural radiance fields with geometry-aware transformers and depth fusion

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Mingqiang Xu,
  • Zhengyao Bai,
  • Chenghao Cao,
  • Zenan Xu,
  • Han Ma,
  • Tao Lin,
  • Qiqin Huang

摘要

Achieving accurate and generalizable 3D reconstruction across diverse scenes remains a major challenge for neural radiance field methods, primarily due to the lack of explicit geometric constraints in existing generalizable frameworks. While current NeRF frameworks achieve impressive photo-realistic renderings, they struggle with scene diversity and geometric fidelity. This paper introduces DeepFusion NeRF (DFN), a Transformer-based framework that integrates monocular depth information to enhance geometric consistency and texture fidelity, specifically in cross-scene novel view synthesis. DFN employs a dual-transformer architecture, comprising a View Transformer and a Ray Transformer, to aggregate multi-view features and decode them along sampled ray points. Monocular depth estimates, provided by an auxiliary network, are fused into the feature representation to enhance geometric consistency without explicit 3D supervision. Our method includes a convolutional gated squeeze–excitation (GSE) module for improved local feature encoding and a composite multi-scale rotary positional encoding scheme for robustness to scale variations. The proposed framework fills the gap left by existing methods that rely on photometric consistency alone, thus improving robustness, especially in occluded or texture-ambiguous regions. Experiments demonstrate DFN’s superior geometric fidelity and competitive rendering quality on both single-scene and cross-scene benchmarks. Codes and models are available at https://github.com/alaneze123/DeepFusion-NeRF-DFN-.