MonoPLFlowNet: Permutohedral Lattice FlowNet for Real-Scale 3D Scene FlowEstimation with Monocular Images

Abstract

Figure 1. Results of our MonoPLFlowNet. With only two consecutive monocular images (left) as input, our MonoPLFlowNet estimates both the depth (middle) and 3D scene flow (right) in real scale. Right shows a zoom-in real-scale scene flow of the two vehicles from side view with the pseudo point cloud generating from the estimated depth map ((middle), where blue points are from frame t, red and green points are blue points translated to frame t+1 by ground truth and estimated 3D scene flow, respectively. The objective is to align green and red points.