NeRF-RERE: a 3D detection network based on 3D NeRF redundancy reduction
Huan Liu, Guimin Jia · 2025
This paper introduces NeRF-RERE, a novel 3D object detection framework leveraging Neural Radiance Fields (NeRF), designed to mitigate computational inefficiencies in complex scene reconstruction. The framework incorporates three key innovations: (1) A density-guided gating module (DGGM) dynamically prunes invalid voxels using density priors; (2) An adaptive depthwise 3D convolution (DyAD-Conv) allocates computing resources via activation intensity perception; (3) A joint optimization approach that integrates NeRF-based scene reconstruction with voxelization to enhance processing efficiency. Experiments on 3D-FRONT and Hypersim datasets demonstrate 4.3% AP and 5.9% AR improvements, while reducing per-epoch training time by >50% under identical hardware, offering a new paradigm for efficient 3D scene understanding.