SCANs: Frequency-Domain Spectrum Channel Attention for LiDAR–Camera Fusion in 3D Object Detection

Abstract

LiDAR–camera fusion has become a widely adopted paradigm for 3D object detection in autonomous driving by effectively leveraging the complementary geometric and semantic information provided by the two sensing modalities. However, most existing fusion methods primarily emphasize cross-modal fusion strategies while relying on conventional feature extraction networks, where global average pooling (GAP) inevitably discards high-frequency information and consequently limits the representation of fine-grained details and long-range contextual dependencies. To address this limitation, we propose Spectrum Channel Attention Networks (SCANs), a frequency-domain-aware feature extraction framework for LiDAR–camera fusion-based 3D object detection. Built upon the Transformer architecture, SCANs incorporates a multi-spectral channel attention mechanism that models feature representations through discrete cosine transform (DCT)-based frequency decomposition. Instead of conventional pooling operations, SCANs employs frequency-aware channel compression to preserve complementary spectral information beyond the lowest-frequency component. Furthermore, we design a learning-based dynamic frequency selection module that adaptively evaluates the task relevance of different spectral components under the 3D detection objective and suppresses redundant or weakly contributing components, enabling more discriminative feature extraction while maintaining computational efficiency. Extensive experiments on the nuScenes, Waymo, and KITTI benchmarks demonstrate that SCANs consistently achieves competitive performance against state-of-the-art methods, validating the effectiveness and generalizability of frequency-domain feature modeling for multi-modal 3D object detection.