Feature Alignment for NeRF-Based 3-D Object Detection
TL;DR
Two lightweight and complementary modules to enhance voxel feature quality for NeRF-based 3D detection with consistent improvements over the NeRF-RPN baseline in both recall and precision are introduced.
Abstract
3D object detection from multi-view images has gained increasing attention as a cost-effective alternative to LiDAR-based methods. However, directly leveraging implicit neural representations such as Neural Radiance Fields (NeRF) for detection faces fundamental challenges, including channel imbalance between RGB and density features and noisy density distributions that degrade localization accuracy. In this paper, we introduce two lightweight and complementary modules to enhance voxel feature quality for NeRF-based 3D detection. First, a geometry-aware fusion module processes appearance and density channels through separate modality-specific branches before recombining them via adaptive fusion, mitigating feature imbalance while amplifying geometric cues. Second, a contour-aware attention mechanism with a density-guided suppression loss reweights voxel features by emphasizing structural boundaries and penalizing background activations, yielding compact and morphology-consistent voxel fields. Extensive experiments on Hypersim, 3D-FRONT, and ScanNet demonstrate consistent improvements over the NeRF-RPN baseline in both recall and precision. Our modules add only 84 parameters and 0.328 GFLOPs, making them readily deployable within existing NeRF-based detection pipelines.