Efficient dense 3D reconstruction using image pairs

This paper presents a dense reconstruction algorithm to obtain a 3D scene with depth information from a pair of images. Accuracy of 3D reconstruction is usually dictated by the correspondence computation between the images. We use Epipolar geometry to improvise the performance of the coarse correspondence matches, and propose to further solve dense correspondence upon these coarse matches through a Epipolar geometry filter and an intensity-based thin plate spline interpolation. We also design systematic evaluation schemes to quantitatively measure the accuracy of this dense 3D reconstruction system.