This is the authors' abstract. We don't add key points for this paper.
To address the performance bottlenecks of unimodal object detection in extreme scenarios such as low light and adverse weather, this paper proposes an efficient Visible-Infrared multimodal object detection framework named CMF-YOLO. The framework aims to achieve high-precision detection and robustness under diverse illumination conditions by deeply fusing the rich textural details of visible imagery with the thermal radiation information of infrared imagery. This approach makes it particularly suitable for complex UAV aerial photography, which is often characterized by small targets and high density. CMF-YOLO organizes its functional components through a coordinated design guided by the frequency characteristics and hierarchical information flow of visible–infrared features. C3k2_CrossFocus incorporates directional slicing and adaptive weighting into the C3k2 feature extraction process, while CrossMambaFusion primarily assigns state-space interaction to low-frequency structural features and retains an independent refinement path for high-frequency details. LMSC-Former replaces quadratic token-to-token self-attention with Global-Gate-guided multi-scale depth-wise convolutional modulation for efficient pixel-wise calibration. Finally, CMMBC-Neck preserves modality-specific shallow details and deep fused semantics through dual-stream direct connections and dense cross-scale shortcuts instead of simple feature concatenation. Extensive experiments on four public datasets—DroneVehicle, VEDAI, M 3 F D , and LLVIP—validate the effectiveness and generalization capability of CMF-YOLO. It achieves an m A P 50 of 86.3 % and m A P 50 - 95 of 65.2 % on the DroneVehicle dataset, an m A P 50 of 75.5 % and m A P 50 - 95 of 45.9 % on VEDAI, and m A P 50 scores of 85.1 % and 96.7 % on the M 3 F D and LLVIP datasets, respectively.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
He et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: