Self-supervised monocular depth estimation based on cross-attention mechanism

Depth information can help robots sense their surroundings. However, it is expensive to acquire depth information through hardware devices such as LiDAR. With the development of technology, self-supervised monocular depth estimation methods have emerged, but there is a lack of estimation of detail information. To improve the prediction of depth, we propose a cross-attention at skip connections of the neural network to reduce the semantic gap of the encoder and decoder. In addition, a convolution-attention hybrid module is used at the decoder minimum scale to enhance the global semantic features. Experimental results on the KITTI dataset show that the proposed method effectively enhances the performance of the depth decoder, giving the network a more detailed depth estimation of the image.