Glass Objects Detection Based on Transformer Encoder-Decoder

Glass objects detection is a relatively new and challenging work in the field of computer vision. This work will also have an impact on other visual tasks, such as depth estimation, instance segmentation and saliency detection. Compared with using high-precision sensors for glass object detection, the vision based detection method is more economical and robust. In this paper, we focus on the task of glass object detection in the field of image segmentation. Traditional semantic segmentation methods mostly used the model based on convolutional neural network to complete the task. On this basis, we introduce the Transformer encoder-decoder architecture to expand the size of receptive field, make full use of different levels of context information and improve the detection effect of glass objects. In this paper, the mixed model of convolutional neural network and Transformer is used to realize the detection of glass objects in the Pytorch environment, and the model is trained and evaluated on the Trans10K-v2 dataset. In the experiment, the input image resolution is 512 × 512, the final pixel accuracy is 93.94%, and the mean intersection over union is 71.32%.