RGB-D-based semantic segmentation method for targets of same shape and different sizes

An RGB-D-based semantic segmentation method for targets of the same shape and different sizes comprises the steps that 1, an RGB-D data set is made, the data set comprises RGB images and depth images, each target in the images is marked, and size categories of the targets are distinguished; 2, constructing a semantic segmentation model network, and training the model to obtain a predicted target category, a bounding box and a pixel position; 3, the point cloud of the target is obtained through the combination of the depth image, the target category, the bounding box and the pixel position; 4, obtaining a scale feature map of the target through the point cloud; 5, inputting the obtained scale feature map into a size classification network, training a size classification model, and obtaining the predicted size of the target; and 6, integrating the results of the steps 2 and 5 to obtain an accurate semantic segmentation result of the target. The method not only can predict categories, bounding boxes and pixel po