Multiscale Efficient Channel Attention for Fusion Lane Line Segmentation

Figure 3

Structure of multimodal multiscale fusion. We use a 3 3 convolution kernel for the image branch and a 5 5 convolution kernel for the point cloud branch.