Fast Robust Large-scale Mapping from Video and Internet Photo ...

More documents

Recommendations

Info

as the GPS/INS reported that the vehicle moved up 10 cm in one example. 4.2. Global Correspondences Since our system creates the camera registration sequentially from the input frames the obtained registrations are always subject to drift. Each small inaccuracy in motion estimation will propagate forward and the absolute positions and motions will be inaccurate. It is therefore necessary to do a global optimization step afterwards to remove the drift. This makes constraints necessary that are capable to remove drift. Such constraints can come from global pose measurements like GPS, but more interesting is to exploit internal consistencies like loops and intersections of the camera path. In this section we will therefore discuss solutions to the challenging task of detecting loops and intersections and using them for global optimization. Registering the camera with respect to the previously estimated path provides an estimate of the accumulated drift error. For robustness the path self-intersection itself can only rely on the views itself and not on the estimated camera motion, which can drift unbounded. Our method determines the path intersection by evaluating the similarity of salient image features in the current frame to all features in all previous views. We found that the SIFT-feature [58] provides a good salient feature for our purposes. For the fast computation of SIFT features, we make use our SiftGPU 2 , which can for example extract SIFT features at 16Hz from 1024 × 768 images on an NVidia GTX280. To enhance the robustness for the video streams we could use local geometry (described in Section 5) to employ our view invariant 2 Available online: http://cs.unc.edu/∼ccwu/siftgpu/ 18
VIP-features [68]. To avoid a computationally prohibitive exhaustive search within all other frames we use the vocabulary tree [69] to find a small set of potentially corresponding views. The vocabulary tree provides a computationally efficient indexing for the set of previous views. It provides a mapping from the feature space to an integer called visual word. The union of the visual words defines the possibly corresponding images through an inverted file. Determining the visual words for the features extracted from the query image requires traversal of the vocabulary tree for each extracted feature in the current view including a number of comparisons for the query feature with the node descriptors. Hereby the features from the query image are handled independently, hence the tree traversal can be performed in parallel for each feature. To optimize performance we employ an in-house CUDA-based approach executed on the GPU for faster determination of the respective visual words. The speed-up induced by the GPU is about 20 on a GeForce GTX280 versus a CPU implementation executed on an Intel Pentium D 3.2Ghz. This allows to perform more descriptor comparisons than in [69], i.e. a deeper tree with a smaller branching factor can be replaced by a shallower tree with a significantly higher number of branches. As pointed out in [70], a broader tree yields to a more uniform and therefore representative sampling of the high-dimensional descriptor space. To further increase the performance of the above described global correspondence search we use our recently proposed index compression method [71]. Since the vocabulary tree only delivers a list of previous views potentially overlapping with the current view we perform a geometric verification of the 19
Page 1 and 2: Fast Robust Large-scale Mapping fro
Page 3 and 4: Figure 1: The left shows an overvie
Page 5 and 6: use statistical models to combine d
Page 7 and 8: a probabilistic way and the final s
Page 9 and 10: streams of multiple cameras mounted
Page 11 and 12: In the case of available GPS data o
Page 13 and 14: 4.1. Camera Pose from Video Our sys
Page 15 and 16: complexity. Our recently proposed A
Page 17: surements we must normalize all of
Page 21 and 22: an active research topic in the com
Page 23 and 24: achieve high computational performa
Page 25 and 26: for each image in the cluster to re
Page 27 and 28: can be added to the 3D model. This
Page 29 and 30: Figure 5: Left: 3D reconstruction o
Page 31 and 32: from 11 images (10 matching, 1 refe
Page 33 and 34: Original Mesh Simplified Mesh Textu
Page 35 and 36: 7. Conclusions In this paper we pre
Page 37 and 38: [12] T. Berg, D. Forsyth, Animals o
Page 39 and 40: [30] Y. Jing, S. Baluja, H. Rowley,
Page 41 and 42: objects from multiple range images,
Page 43 and 44: [63] M. Fischler, R. Bolles, Random
Page 45 and 46: Marquardt algorithm, Tech. Rep. 340

Fast Robust Large-scale Mapping from Video and Internet Photo ...

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?