PHKD: PARALLEL HIERARCHICAL KEYFRAME DECIMATION OF VIDEO FOR 3D RECONSTRUCTION

In automatic Structure from motion (SFM), the keyframe selection or decimation is essential, which makes the scene reconstruction more efficient and precise. However, there is a gap in efficiency and robustness between current keyframe selection procedure and the practical use. In this paper, we propose a parallel hierarchical keyframe selection framework for 3D reconstruction with video sequences. First, it operates in real time by introducing an optimized procedure to store and operate keyframes. Then it manages the practical issues like object tracking lost, repetitive motion and redundancy with the lost-recover and fine selection procedure. Finally, we evaluate the framework on a video sequence datasets captured by UAVs and mobile devices. Our experimental result illustrates that both the execution time, robustness to the input video and model quality are improved compared with the general keyframe selection approaches for 3D reconstruction.