The deformable convolution module introduces an offset for the sampling factors, as illustrated in Fig. two. The very best portion generates the index offset by processing the input characteristic map as a result of a regular convolution layer, while The underside component convolves the input feature map Along with the corresponding kernel to pro