3D Structure from a Single Calibrated View Using Distance Constraints
Rubin Gong, Gang Xu, Regular Member · Medical Entomology and Zoology · 2003
SUMMARY We propose a new method to recover scene points from asingle calibrated view using a subset of distances among the points. Thispaperfirstintroducestheproblemanditsrelationship withtheperspective n point problem. Then the number of distances required to uniquely recoverscene points are explored. The result is then developed into a practicalvision algorithm to calculate the initial points’ coordinates using distanceconstraints. Finally SQP (Sequential Quadratic Programming) is used tooptimize the initial estimations. It can minimize a cost function defined asthe sum of squared reprojection errors while keeping the specified distanceconstraints strictly satisfied. Both simulation data and real scene imageshave been used to test the proposed method, and good results have beenobtained. key words: perspective n point problem, single view reconstruction, sensi-tivity analysis, sequential quadratic programming 1. Introduction Traditional image-based 3D reconstruction methods usemultiple views to extract 3D geometry. However, when re-constructing scene structures with only one photograph orretrieving dynamic scene properties shot by one video cam-era, only a single view exists. To recover 3D scene struc-ture from a single view, 3D scene constraints need to beutilized[1]. A variety of 3D scene geometry constraintshavebeenidentifiedandusedincomputervisionalgorithms.Traditional constraints have used vanishing points, planesand geometric invariants constraints[2]–[4]. Some workhas used the constraint of multiple objects of the same typein the scene separated by a simple translation[5]. How-ever,thesegeometryconstraintsarelimitedtocertainmodeltypes,forexample, scenescomposedofplanesorothersim-ple primitives.The 3D scene distance constraints are used to recover3D models from a single view in this paper. Since 3D scenedistance constraints exist in almost every kind of scenes,the proposed technique can be applied in a wide variety ofscenes that contain points, lines, and planaror freeform sur-faces. In multiple view cases, 3D scene distance constrainthas been applied in motion capture applications to upgradeaffine reconstruction to a metric one[6]. In single view re-construction applications, the 3D scene distance constraintshave not been effectively utilized so far.The closest work to ours is the traditional perspective