Neural Prior for Trajectory Estimation
Supplementary Material
Paper ID: 9911

Visual results for trajectory estimation on Argoverse dataset

We provide more video results for trajecoty estimation on the Argoverse dataset. As mentioned in the main paper, we used a sequence of 25 lidar frames sampled at 10 Hz, and estimated the dense trajectory for each point. The following videos show the estimated trajectories at each frame. Our method directly ouputs the full trajectories, while NSFP [1] used forward Euler integration, and FlowStep3D [2] used K-nearest neighbor integration. Colored points represent the point cloud, and the gradient gray lines denote the estimated trajecory. The light gray lines are trajectories in older frames, and the darker gray lines indicate the most recent motions.

NTP (Ours) NSFP (Euler Int.) Flowstep3D (KNN Int.)

Failure case

When the point cloud is too sparse, the trajectory estimation across a long sequence might suffer from lacking adequate corresponding query points. We provide a failure case below to show how the sparse point affects our method.

Point cloud densification

We provide more visual results on the Argoverse dataset. For each example, the upper row shows the densified depths compared to the sparse depth, and the bottom row shows the densified point clouds compared to the original sparse point cloud.

Sparse reference frame NTP (Ours) NSFP (Euler Int.) Flowstep3D (KNN Int.)

Here are videos showing how the densified point clouds across the long sequence compared to the original sparse point clouds.

Densified point cloud using NTP (Ours) Original sparse point cloud

Dense non-rigid structure from motion

We visualize our dense NRSfM result on the synthetic face traj. B sequence. Our reconstructed mesh is colored to visualize the trajectory code field.

NTP (Ours)


[1] Li, Xueqian, Jhony Kaesemodel Pontes, and Simon Lucey. "Neural Scene Flow Prior." Thirty-Fifth Conference on Neural Information Processing Systems. 2021.
[2] Kittenplon, Yair, Yonina C. Eldar, and Dan Raviv. "FlowStep3D: Model Unrolling for Self-Supervised Scene Flow Estimation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.