What is dynamic convolution and how does it differ from standard convolution?

Dynamic convolution generates input-dependent kernels instead of using fixed convolution filters.

Standard convolution uses fixed learned kernels for all inputs. Dynamic convolution introduces a mechanism where convolutional kernels are conditioned on the input feature map. A lightweight network (often attention-based) generates weights to combine multiple candidate kernels dynamically. This allows the model to adapt its receptive behavior depending on the image content, improving performance on diverse scenes and reducing redundancy in learned filters.