Any Motion Detector: Learning Class-agnostic Scene Dynamics from a Sequence of LiDAR Point Clouds

Scalable Scene Flow from Point Clouds in the Real World

“…We observe that mean L 2 error increases substantially when ego motion is not compensated for across all object types and across moving and stationary objects. This is also consistent with previous works [16]. We also ran a similar experiment where the model consumes non ego motion compensated point clouds, but instead subtracts ego motion from the predicted flow during training and evaluation.…”

“…We argue that this is more realistic for AV applications in which ego motion is available from IMU/GPS sensors [49]. Furthermore, having a consistent coordinate frame for both input frames lessens the burden on a model to correspond moving objects between frames [16] as explored in Appendix B.…”

“…The average scene from the Waymo Open Dataset consists of 177K points (Table 2), even though most models [28,56,21,54,29] were designed to train with 8,192 points (16,384 points in [29]). These design choices result in favoring algorithms that scale poorly to O(100K) regimes.…”