EN-Transformer: an improved dual-transformer absolute pose estimation method based on position and orientation

Camera pose estimation refers to the method of accurately estimating the three-dimensional position coordinates and threedimensional angular orientation of the camera based on a given image. It is a key technology for applications such as autonomous driving, virtual reality, and instant localization. Currently, camera absolute pose regression methods based on deep learning quickly predict the camera pose of the query image through a trained end-to-end network model with a single forward propagation. However, in environments with target occlusion, lighting changes, dynamic blurring, and sparse or repetitive scene textures, the robustness and generalization of the network are insufficient, leading to poor localization accuracy. To address these issues, we use MS-Transformer as the backbone network to construct a more efficient ENTransformer model for multi-scene absolute pose regression. Specifically, we designed a multi-scale parallel fusion module to obtain features with different receptive fields, enhancing the network's performance in handling complex scenes. We then designed and introduced External Attention to synergize channel and external attention, allowing better focus on extracting sparse features, reducing the model's reliance on a single channel, and improving the performance of regression tasks. By introducing the MUSE module, we combine multi-scale and self-attention. Finally, by leveraging the characteristics of different modules, we design different transformers to focus on specific tasks, handling position translation and direction rotation information separately. To assess the model's effectiveness, experiments were carried out using the widely adopted 7Scenes dataset. The results indicate that the detection accuracy on this dataset has enhanced by roughly 5%.