Semantic Segmentation of Transparent Objects Based on 3D Weights

Due to the fact that transparent objects are optically transparent by themselves and at the same time lack obvious surface features. Simply using Convolutional Neural Network or Transformer Network can easily lead to problems in the segmentation process, in this paper, we combine both of them while adding 3D weights to make up for the shortcomings of the lack of surface features of transparent objects. Experimental results on the Trans10K-v2 dataset show that the improved method in this paper can effectively improve the average intersection ratio and accuracy of semantic segmentation compared with other semantic segmentation methods.