A Method of Hierarchical Image Retrieval for Real-Time Photogrammetry Based on Multiple Features

In near real-time photogrammetry, the first step in processing each new added image is determining the most relevant image in pre-sequence unordered images quickly and exactly, which is pivotal for accurate image matching and 3D reconstruction. This paper presents a hierarchical image retrieval algorithm based on multiple features and details the choice for representation of multiple features which is critical to the improvement of accuracy of this algorithm. First, we represent global features using AlexNet-FC7(fully connected layers) or ResNet101-Pool5(pooling layers) and local features using SIFT (scale-invariant feature transform) in two parallel threads with support of GPU (Graphics Processing Unit). Next, we obtain candidates based on cosine similarities between global features of each pre-sequence image and new added image. Finally, we determine the most relevant image from those candidates according to feature matching results for each candidate and new added image. The experimental results confirm that the second step is rather fast and the third step is necessary to tackle the problem that global features cannot distinguish objects from the same class. The total time our algorithm takes is about 83.6ms for determining the most relevant image in 5063 pre-sequence unordered images of size
$1024\times 768$
, which outperforms exhaustive pairwise matching, Bag of Words and multi-vocabulary trees. Accuracy of our algorithm also perform better than the state-of-the-art methods on three benchmark datasets. SIFT matching results obtained in the third step after eliminating mismatches with RANSAC (Random Sample Consensus) can also be used for high-precision incremental SFM (Structure from Motion) reconstruction.