Stereophotogrammetry follows the principle of human visual perception to interpret the 3D information of an object from a digital camera.
In order to achieve this, stereo vision cameras are designed to mimic human eyes by capturing images from two lenses focusing on the same object but having different and partially overlapping viewpoints. This allows the cameras to capture the orientation, depth and other geometrical information that can be further processed mathematically to reconstruct a 3D view of the object. [1].
Fig. 1 Two cameras viewing the same 3D point of an object on different 2D locations [3]
Depth of an object is highly important in reconstructing a 3D shape as this gives a new dimension to a 2D image because building a 3D shape with a single 2D RGB image is challenging, and it has its limitations. To solve this problem the depth information has been incorporated in the reconstruction process as this has been proven to reduce the reconstruction error. [1]
Depth estimation using a single image can be done either using monocular depth estimation algorithms to get an estimate of the depth or using RGB-D cameras as the depth is recorded by the depth camera. Estimating depth from stereo cameras is comparatively more accurate than monocular depth estimation because of the two viewpoints.[1]
Before calculating the depth, the stereo cameras are calibrated, images are rectified and two images of the same object with corresponding feature points are triangulated to extract the depth values mathematically. Determining the placement of the cameras is also essential in reducing the 3D reconstruction error and accurate depth calculation.
By using the depth maps and images, a 3D model can be constructed. Deep learning methods can also be used to recover a 3D representation of the object as it explores the dense corresponding feature points of the two images.[1] The predicted 3D points form a point cloud and are further enhanced by surface reconstruction and texture mapping to get a real 3D view. Stereo vision has a diverse range of applications but there are more challenges to overcome such as occlusion, illumination, resolution and noise.
Bibliography
[1] Haozhe Xie, Hongxun Yao, Shangchen Zhou, Shengping Zhang, Xiaoshuai Sun, Wenxiu Sun.Toward 3D Object Reconstruction from Stereo Images, doi ={10.48550/arXiv.1910.08223}
[2] do Phuong, Binh & Nguyen, Quoc Chi. A Review of Stereo-Photogrammetry Method for 3-D Reconstruction in Computer Vision. 138-143. 10.1109/ISCIT.2019.8905144, doi = {10.1109/ISCIT.2019.8905144}, 2019
[3] Borangiu, Theodor & Dumitrache, Alexandru. Robot Arms with 3D Vision Capabilities. Doi = {10.5772/9668}, 2010