Improvement of Light Field Feature
Xi Chen, Murat Tahtalı, Matthew Garratt, Mark R. Pickering · 2019
The recent development of light field feature detectors has brought light field imaging into many 3D reconstruction applications. Compared with the traditional 2D, higher dimensional feature descriptors reveal more angular information, but with significantly decreased efficiency in structure from motion (SfM) applications. To improve the efficiency of feature extraction, a new 4D light field feature detector and descriptor (LiFF) was invented by Dansereau et al. [1]. LiFF is scale and rotation invariant with refocusing ability, which makes the features more robust to partial occlusions. Moreover, structure from motion software, such as Colmap, can work with this a prior knowledge to benefit the incremental reconstruction. LiFF uses joint search technology for scale and slope space, and feature extraction speed is increased by more than ten times. However, the efficiency and accuracy of LiFF can be further improved by optimizing the scale and slope space. In this work, we add a rough slope estimation into the feature extraction pipeline, which can further improve the speed of LiFF.