PT3 Prototype-Guided Pretraining for mmWave Human Point Clouds
October 30, 2026
PT3 uses a prototype-guided self-supervised framework to learn stable human representations from sparse and noisy mmWave radar data. By representing spatial states as compositions of shared prototypes, the encoder enables downstream tasks like pose estimation and motion prediction without retraining.
HOW THIS AFFECTS YOU
●
builderThis framework offers a more robust backbone for building radar-based human sensing applications.
●
researcherThe use of spatial prototypes provides a method to stabilize representations against radar-specific perturbations.