Simulated and phantom colon data for 6-dof camera pose estimation

Since narrow imaging and texture scarceness issues results to high missing rate  for gastrointestinal disease in clinical colonoscopy, the pose and depth estimation is required to improve the quality of diagnosis.    Some unsupervised methods have been proposed for this purpose. However, these methods often suffer from the performance degradation for a long and curved trajectory because of the structural constraints. In this paper, we propose an innovative key points encoding Transformer (KPEformer) for pose estimation in long and curved colonoscopy trajectory. Specifically, a light weight convolution and attention module is proposed to encode the features of latent variables for key points. Two loss functions for confidence and repeatability of key points are designed to enhance the reliabilty of output information. In addition, we propose a more adequate colon dataset which contains synthetic    and phantom to remedy the deficiency of existing dataset. Extensive experiments on our proposed synthetic dataset and phantom dataset demonstrate the state-of-the-art performance of our method.