| 英文摘要 | With the popularity and easy access of high resolution images in recent years,many computer vision problems become computationally prohibitive, and GPU based acceleration is considered the most effective way to tackle the problem. This thesis is focused on GPU based scene reconstruction and object modeling, in particular, the involved key steps, such as feature detection, matching, and final optimization, and the main contributions include: (1) The image feature detection and the feature matching. We take two widely used algorithms as representative examples to evaluate the performance of the GPU implementation, that is, the Scale-invariant feature transform (SIFT) for feature detection and the KD-Tree traversal for feature matching. We present the design, implementation, and evaluation of these two algorithms on GPU for high-resolution image datasets. Comparisons are performed on the implementation on CPU and GPU using classical Oxford’s database as well as our own high-resolution image datasets. Our results show that around 95% of the extracted features are nearly the same, and the repeatability score and matching score are similar under the GPU and CPU implementations under various image changes, such as viewpoint change, scale change, rotation, bluring and illumination variation. The runtime speedup of the GPU implementation for the SIFT detection is about 13x faster than their CPU counterpart, and that for the KD-Tree traversal is about 43x. In sum, our results show that the GPU implementations for feature detection and matching are as good as the CPU ones, and can be safely used in real applications. (2) GPU Based Bundle Adjustment(BA). BA is a crucial step in 3D reconstruction but time consuming. In this thesis, a special class of BA problems are investigated, where the reconstructed 3D points are much numerous than the camera parameters, namely Massive-Points BA (MPBA) problems. We present the design and implementation of a new bundle adjustment algorithm for efficiently solving the MPBA problems. The hardware parallelism, the multi-core CPUs as well as GPUs, is explored. By careful memory-usage design, the graphic-memory limitation is effectively alleviated. Several modern acceleration strategies for bundle adjustment are explored and compared. By using several high-resolution image datasets, we generate a variety of MPBA problems, with which the performance of five bundle adjustment algorithms are evaluated. The experimental results show th... |
修改评论