A2D2C: Adaptive attention-driven dynamic convolution for local feature adaptation

Key Points

• To introduce A2D2C, a novel attention-driven dynamic convolution method for local feature adaptation.
• Utilized multi-point random sampling to efficiently route k base kernels.
• Implemented A2D2C+ for effective kernel fusion, reducing redundancy in model computations.
• Demonstrated consistent performance improvements on ImageNet, CIFAR-100, and COCO datasets.