Feature Descriptor Learning Based on Sparse Feature Matching

The 3D structure reconstruction of endoscopic images is critical for endoscopic-guided surgical navigation systems. Besides, point correspondence estimation of endoscopic images is a critical step to realize 3D structure reconstruction. However, stable and dense matching points are difficult to obtain. We propose a feature descriptor learning method based on sparse feature matching to overcome this limitation. A few matching points were produced for supervised network training by adopting a classical feature matching method, where weight adaptive technique was utilized to mitigate the influence of mismatched points. An end-to-end network architecture was constructed to map endoscopic images to feature descriptor maps and avoid checkerboard artifacts. The proposed method was evaluated on the Stereo Correspondence and Reconstruction of Endoscopic Data and Endoscopic Simultaneous Localization and Mapping datasets. Results showed that our method was able to extract feature descriptors from endoscopic images effectively and simultaneously obtained denser and more accurate matching points.