SFM1 estimates the locations of setof 3D points from a series of images that gives a sparse set of correspondencesbetween image features. The 3D geometry (structure) and camera pose (motion)are estimated simultaneously. It is commonly known as structure from motion.
Photogrammetry3 is the technology of extractinginformation such as pixel positions from images. Photogrammetry has beensuccessfully employed in a wide range of industries. The main application ofphotogrammetry3 is to generate 3D models out of the images taken from anobject, and also used for many different purposes. Limitations such asreflective surfaces, transparency issues, complicated shapes with smallcomponents, etc.
are some of the major challenges to be improved.I. related work A. Feature ExtractionFeature extraction and matching isthe foundation of SFM1 method, its goal is to fmd and correct matching keypoints of the same object or scene in the two images. Harris8 proposes one ofthe earliest corner detectors. Another widely used keypoints at the moment isSURF. It has similar matching performances as SIFT7, but it’s much faster.However, the dimensionality of the feature vector is still too high forlarge-scale 3D reconstruction.
BRISK9 makes use of an easily configurablecircular sampling pattern from which it computes brightness comparisons to forma binary descriptor string. Comparing with the random sampling mode in BRISK. B.
Image MatchingTime complexity of image matching isdecided by two aspects: one is the time complexity of similarity comparison, theother is the time complexity of search. In order to reduce the time complexity,we often creates a kd tree to organize the feature sets and using KNN(K-nearestneighbor) algorithm to speed up the matching process. K-Means algorithm is usedto quantizing the image features and Brute-Force algorithm to compare featuresof images. C. Fundamental MatrixThe estimation of fundamental matrixis commonly using RANSAC9. The fundamental matrix between two images isknown, it can realize the projective reconstruction. D.
Camera CalibrationCamera calibration target is todetermine the relationship between image coordinate and world coordinate. Theaccuracy of calibration directly affects the results of the subsequentreconstruction. Traditional camera calibration such as DLT(Direct Linear Transformation),can achieve high accuracy, but usually algorithm is more complex depending onthe high precision calibration block. The SFM1 technology relies on thefundamental matrix F which plays a very important role in the process ofrecovering certain information about camera intrinsic. E. Bundle AdjustmentBundle Adjustment (BA)11 is the keytechnology of SFM, the most accurate way to recover structure and motion is to performrobust non-linear minimization of the re-projection errors called bundleadjustment.
Sparse bundle adjustment12 (SBA) is an available high- qualityalgorithm based on incremental standard equation.