HINet: Hierarchical Point Cloud Frame Interpolation Network

Abstract

Intelligent agents utilize various sensors such as LiDAR, cameras to perceive the surroundings. However, the frame rate difference among sensors seriously affects both safety and efficiency of intelligent agents. Recently some research concerning point cloud frame interpolation is conducted to solve the frame rate inconsistency problem by interpolating low frame rate point cloud sequences up to high frame rate ones. To improve the performance of current state-of-the-art method, we come up with a novel Hierarchical Point Cloud Frame Interpolation Network (HINet). By proposed hierarchical warping module, coarse intermediate frames are generated hierarchically to reach closer toward the target position. Besides, we propose spatial aware fusion strategy to hierarchically restore local geometric distribution by attention mechanism and positional offset. Finally, hierarchical supervision module is applied to efficiently train the HINet in two stages, guaranteeing the quality of predicted intermediate frames. We employ HINet in a large outdoor autonomous driving dataset and provide convincing qualitative and quantitative evaluation results.