is RFB position sensitive?

[birdwcp]

        loc_layers += [nn.Conv2d(v.out_channels, cfg[i]
                             * 4,kernel_size=3, padding=1)]
        conf_layers += [nn.Conv2d(v.out_channels, cfg[i]
                              * num_classes, kernel_size=3, padding=1)]

the kernel_size is 3 in the the loc_layers and conf_layers. From My Understanding，to make the anchor boxes more consistent with the Receptive field, should the kernel_size be set 1? For example，in the 5x5 feature map, the size of anchor boxes is about 3x3 , the branch in RFB(rate=1) tailed with 3x3 kernel conf_layer makes RF to be about 5x5 in the 5x5 feature map. if set conf_layer kernel to be 1x1 , the RF will be 3x3 , similar with anchor boxes .