Scene Reconstruction From Internet Image Collections Base On Structure From Motion

The main content of this paper is about scene reconstruction of internet image collections basing on SfM,and a more popular method to solve this problem is incremental reconstruction which includes matching stage and reconstruction stage.Inspired by the special features of internet image collections,I proposed a method to remove the redundant information from the collections to raise efficiency at matching stage,and taking account of image scale,I proposed a reconstruction method called From Global to Local reconstruction(G2L)to improve global accuracy and efficiency at the second stage.At matching stage,basing on GraphMatch,I presented ThinMatch,which is an efficient method for matching of large-scale images.The main ideas of this method are as follows: because large-scale images are usually from internet,there must be lots of redundant information,such as redundant images which are unrelated to the main object,redundant features which are unrelated content in an image,such as people,animals and trees,redundant compute which we use to match an image with those which are related to the object but do not match this image.For these three types of redundancies,ThinMatch uses pedestrian detection to filter redundant content in an image,uses the last match rate to filter unrelated images,uses top n large scale features to filter unmatched images.In the experiment,I compare my method to brute-force matching approach and GraphMatch(state-of-the-art),it is shown that this method effectively raises the matching speed and guarantees a certain quality.Especially at the stage of reducing compute redundancy,comparing to GraphMatch,it doubles the matching speed and remains about 94% of the number of registered cameras averagely.At reconstruction stage,I proposed a method called From Global to Local reconstruction(G2L)which effectively improved two issues in the process of reconstruction.Firstly,for now,as far as I know none of the proposed methods take image scale into account.Some images which were shot from a long distance or by wide-angle lens capture the structure of the scene,while some images which may be shot from a short distance or by teleimage lens capture the details of the scene.When adds a new image to the model in an incremental method,scales were not taken into account,and those images with different scales were put into a same optimization system,which may cause a global shift because of local optimization.Secondly,in the process of reconstruction of large-scale images,with the growth of the number of images,the parameters in the optimization system will become larger,which has a strong impact on the efficiency of incremental methods because this method need to run optimization frequently.Our method performed as the following steps.Firstly,calculate the image scale ratio of image pairs from scale ratio of features in the image pair,and then calculate the image scale from image scale ratio.Secondly,build a scale pyramid basing on the image scale,and build a configuration pyramid basing on the scale pyramid.Finally,reconstruct the scene level by level according to the configuration pyramid.Fix the parameter in the model once all of the images in a level are reconstructed which can avoid the effect from the lower level of images,decrease the number of parameters in the optimization system,and hence improve the efficiency of reconstruction.The experiment shows that this method efficiently reduces the time for reconstruction and the global shift of the model.