NeRF-FCM: Feature Calibration Mechanisms for NeRF-based 3D Object Detection

With the fast development of 3D vision, 3D object detection based on posed RGB images has become increasingly popular and attracted significant attention from researchers in recent years. Given the remarkable performance of Neural Radiance Field (NeRF) in modeling 3D scenes, recent 3D detection methods utilizing posed RGB images generated by NeRF models have achieved promising results. However, NeRF-based models often suffer from poor generalization and are prone to generating inconsistent image content for unseen views, which inevitably degrades the performance of existing NeRF-based 3D detectors. In this paper, we propose an effective feature calibration method to enhance the performance of 3D detection models based on posed RGB images produced by NeRF models. Specifically, our proposed method efficiently recalibrates the 3D features extracted from the backbone network, and adaptively computes the weights for fusion based on the statistical properties of the features. Experiments show that our method significantly outperforms the baseline model, achieving improvement of +8.6 [email protected], +5.5 [email protected], and +5.1 [email protected] on the Hypersim, 3D-FRONT, and ScanNet benchmarks, respectively, with anchor-free heads. Particularly, compared with the baseline model, our method can more accurately predict 3D bounding boxes in 3D space, even when objects are poorly reconstructed by NeRF while keeping low computational costs with a minimal increase in model complexity.