Dense 3D Reconstucion COMS 6733 3D Photography Final Project

Current 3D reconstruction methods using stereo vision can produce good results, but they tend to yield very sparse outputs and suffer from several challenges. One of them is the well-known correspondence problem, whereby similar features in different images are often too ambiguous to match reliably. The problem is accentuated when the object being scanned is textureless. Under these circumstances, methods that are widely used for correspondence determination, such as SIFT, normalized cross-correlation of corner features, and histogram descriptor approaches, will not be able to produce reliable results for matching stereo features for 3D triangulation. Our project consisted of two different approaches, each trying to alleviate these issues. One is based on photometric stereo. The second is based on a two-camera stereo system with a finely-controlled laser pointer. The former did not yield the expected results, but it supplied us with useful information to succeed with our second approach, which turned out to be much more successful. Both methods tried to take advantage of the Staubli robot arm for its ability to achieve fine-tuned positional accuracy. In the first case, we used it to position the light source, and in the second case we used it to control the motion of the laser pointer. In this report, we will discuss the procedures for both approaches, analyze the influential factors for the failure of the first approach, and show the experimental results of the second approach and discuss some areas for improvement.