关键词: 摄影测量, 三维重建, 运动恢复结构, 深度特征, 卷积神经网络

Abstract: Reliable feature matching plays an essential role in SfM (structure from motion) for UAV (unmanned aerial vehicle) images. Recently, deep learning-based methods have been used for feature detection and matching, which outperforms traditional handcrafted methods, e.g., SIFT, on benchmark datasets. However, few studies have reported their performance on UAV images as these models are trained and tested using internet photos. By using UAV datasets with varying features, this study evaluated both handcrafted and learned methods in terms of feature matching and SfM-based image orientation. The experimental results show that even with the pretrained public-available models, more accurate and complete feature matching can be obtained through the combination of high-precision localization of handcrafted detectors and the high representation ability of learned descriptors, which has competitive or better performance in SfM-based image orientation when compared with SIFT-like handcrafted methods.

Key words: photogrammetry, 3D reconstruction, structure from motion, learned feature, convolutional neural network