SynFlow: Scaling Up LiDAR Scene Flow Estimation with Synthetic Data

SynFlow addresses a major bottleneck in autonomous driving: the difficulty of obtaining ground-truth labels for non-rigid 3D motion (scene flow). While self-supervised methods like NSFP (Neural Scene Flow Prior) or FastNSFP have gained traction, they suffer from noisy signals on real-world data. SynFlow's 'paradigm shift' towards pure synthetic scaling is a logical progression seen in other computer vision domains (like optical flow with FlyingChairs), but it lacks a moat beyond the specific simulation parameters and model weights. With 0 stars and only 4 forks, it is currently a fresh research artifact without community gravity. The defensibility is low because the competitive advantage relies on the quality of the synthetic data generation pipeline, which frontier autonomous driving labs (Waymo, Tesla, Zoox) already possess internally. The project faces high platform domination risk as these labs are unlikely to adopt an open-source motion prior when they can generate proprietary simulation data tailored to their specific sensor suites.