This paper proposes a method for reconstructing 3D detailed structures of internal organs such as gastric wall from
endoscopic video sequences. The proposed method consists of the four major steps: Feature-point-based 3D
reconstruction, 3D point cloud stitching, dense point cloud creation and Poisson surface reconstruction. Before the first
step, we partition one video sequence into groups, where each group consists of two successive frames (image pairs), and
each pair in each group contains one overlapping part, which is used as a stitching region. Fist, the 3D point cloud of
each group is reconstructed by utilizing structure from motion (SFM). Secondly, a scheme based on SIFT features
registers and stitches the obtained 3D point clouds, by estimating the transformation matrix of the overlapping part
between different groups with high accuracy and efficiency. Thirdly, we select the most robust SIFT feature points as the
seed points, and then obtain the dense point cloud from sparse point cloud via a depth testing method presented by
Furukawa. Finally, by utilizing Poisson surface reconstruction, polygonal patches for the internal organs are obtained.
Experimental results demonstrate that the proposed method achieves a high accuracy and efficiency for 3D
reconstruction of gastric surface from an endoscopic video sequence.