Towards Robust Homography Estimation for Forward-Motion Panorama for Multi-camera Wireless Capsule Endoscopy Videos

A noteworthy challenge in Wireless Capsule Endoscopy is the significant time and expertise required by medical professionals to accurately interpret its videos and detect anomalies. The combination of a large number of frames to analyze within each video coupled with poor image quality contributes to low lesion detection rates. To address these issues, our study explores a methodology for the construction of local forward and/or backward-motion panoramic overviews from videos rendered from a synthetic tubular model created in Blender. This approach aims to consolidate essential information from multiple frames for the purpose of lesion detection and localization. The mosaicing process is explored by computing global homographies between sequential frames and three methods for homography estimation are assessed. This study compares the resulting local panoramas obtained with our method with those produced using a deep neural network approach, utilizing three different models for image quality assessment.