Robust and efficient 3D perception is critical for autonomous vehicles operating in complex environments. Multi-sensor fusion, such as Camera+LiDAR, Camera+Radar, or all three modalities, significantly enhances scene understanding, However, most existing frameworks fuse data at a fixed stage, categorized as early fusion (raw data level), mid fusion (intermediate feature level), or late fusion (detection output level), neglecting semantic consistency across modalities. This static strategy may result in performance degradation or unnecessary computation under sensor misalignment or noise. In this work, we propose FDSNet (Feature Disagreement Score Network), a dynamic fusion framework that adaptively selects the fusion stage based on measured semantic consistency across sensor modalities. Each sensor stream (Camera, LiDAR, and Radar) independently extracts mid-level features, which are then transformed into a common Bird’s Eye View (BEV) representation, ensuring spatial alignment across modalities. To assess agreement, a Feature Disagreement Score (FDS) is computed at each BEV location by measuring statistical deviation across modality features. These local scores are aggregated into a global FDS value, which is compared against threshold to determine the fusion strategy. A low FDS, indicating strong semantic consistency across modalities, triggers mid-level fusion for computational efficiency, whereas a high FDS value activates late fusion to preserve detection robustness under cross-modal disagreement. We evaluate FDSNet on the nuScenes dataset across multiple configurations: Camera+Radar, Camera+LiDAR, and Camera+Radar+LiDAR. Experimental results demonstrate that FDSNet achieves consistent improvements over recent multimodal baselines, with gains of up to +3.0% in NDS and +2.6% in mAP on the validation set, and +2.1% in NDS and +1.6% in mAP on the test set, highlighting that dynamic stage selection provides both robustness and quantifiable advantages over static fusion strategies.

Object detection is a fundamental task in computer vision. It has two primary types: one-stage detectors known for their high speed and efficiency, and two-stage detectors, which offer higher accuracy but are often slower due to their complex architecture. Balancing these two aspects has been a significant challenge in the field. RetinaNet, a premier single-stage object detector, is renowned for its remarkable balance between speed and accuracy. Its success is largely due to the groundbreaking focal loss function, which adeptly addresses the issue of class imbalance prevalent in object detection tasks. This innovative approach significantly enhances detection accuracy while maintaining high speed, making RetinaNet an ideal choice for a wide range of real-world applications. However, its performance decreases when applied to datasets containing objects with unique characteristics, such as objects with elongated or squat shapes. In such cases, the default anchor parameters may not fully meet the requirements of these specialized objects. To overcome this limitation, we present an enhancement to the RetinaNet model to improve its ability to handle variations in objects across different domains. Specifically, we propose an optimization algorithm based on Differential Evolution (DE) that adjusts anchor scales and ratios while determining the most appropriate number of these parameters for each dataset based on the annotated data. Through extensive experiments on datasets spanning diverse domains such as the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI), the Unconstrained Face Detection Dataset (UFDD), the TomatoPlantFactoryDataset, and the widely used Common Objects in Context (COCO) 2017 benchmark, we demonstrate that our proposed method significantly outperforms both the original RetinaNet and anchor-free methods by a considerable margin.