Research On Acceleration Technology Of Large-scale Scene3D Reconstruction

Image based modeling is a research hotspot of computer vision during past decades. With the rapid rising requirement of movies, games, virtual display and culture heritages protection, we are observing a significant breakthrough in image based three dimensional reconstruction technologies. As for large-scale scene reconstruction, image based methods are ideal choice since traditional3D modeling software is interaction intensive and structural light or laser scan based technology is so expensive which makes it impractical for large scene. The core research of this paper lies on image based reconstruction technology for large-scale archaeological sites and its acceleration strategy.My research roots in archaeological practice of large-scale historical sites reconstruction. Our technical route as follows:we detect image features for each photo and matching these features between different images first followed by calculating camera pose using self-calibration technique based on feature correspondences. Next step we divide the whole image set into several clusters and calculate dense point cloud geometry of scene for each cluster using multi-view stereo technique. Finally, we can get the final model by merging these could point together followed by some simple manual processing.My research work focuses on exploring multi-view reconstruction technology and its acceleration strategy for large-scale scene reconstruction. The main contribution of this paper includes following three aspects:First, we design and implement a selective feature matching solution using visual vocabulary strategy. For large sites reconstruction the image sets could be extremely large and process feature matching between all image pairs will become unrealistic. So we use bag of words model based on visual words to screen image pair which probably share common scene elements to get rid of full-match bottleneck. Second, we propose a new SIFT feature matching algorithm based on clustering technology. The basic idea is we divide all features of image set into many subspace first and then process feature matching operation within each subspace. Reconstruction practice shows performance of the new method outweighs standard SIFT matching algorithm for large sites reconstruction. In the third part we attempted to improve the performance of computation-intensive module with distributed computing strategy, designed and implemented an initial distributed three dimensional reconstruction system. Engineering practice shows that technique routes suitable for large-scale sites modeling and distributed acceleration strategy is a promising direction which deserve further trial.