mmWave radar micro-Doppler becomes fittable: 4DPS turns a real body capture into trainable radar data

Human micro-Doppler from millimeter-wave radar can now be inverted from a real capture, rather than only simulated forward or sampled by a generative model. Earlier radar inverse renderers reconstructed static scenes only, evaluating one phase per point per frame, so they had no slow-time axis along which Doppler is measured and could neither define velocity nor fit a moving body. On HuPR, fitting raises the clutter-removed range–azimuth correlation from 0.270 for the camera-initialized render to 0.666; the fitted renders reach median micro-Doppler correlations of 0.988 on HuPR and 0.966 on RT-Pose, against 0.671 and 0.394 for the forward simulator RF-Genesis driven by the same motion.

The renderer outputs the raw complex return a real radar records, so one fitted scene passes through the same digital signal processing as real data and yields any representation a downstream model reads. Earlier static inverse renderers had no Doppler axis, and learned generators left the scene implicit in network weights, with nothing explicit to fit or edit. The scene is initialized from the capture's synchronized camera, each point carries an ITU-R P.2040 material, a normal and a velocity, and moving points are splatted along range and Doppler in closed form, depositing into seven Doppler bins per point; fitting covers 235 HuPR and 100 RT-Pose sequences at about 5.7 and 12.9 minutes per sequence on one H100.

Pose models trained only on rendered data work on real radar data, showing that the renders carry the information the task needs. Models trained on renders from the forward simulator RF-Genesis, driven by the same motion, transfer almost nothing to real data. On RT-Pose, a model trained on 4DPS renders alone reaches 21.86 cm MPJPE on real data, against 16.67 cm for real training and 127.59 cm for RF-Genesis renders; on HuPR the three arms reach 38.4 AP, 65.6 AP and 1.8 AP.

The demonstrated scope is two public human-sensing datasets, HuPR and RT-Pose, that retain slow time and ship a synchronized camera, covering two different radar array geometries; transport is single-bounce, the background is held static, and one person is modeled per sequence. The authors name scoring edited scenes against real recordings as the next step.

Velocity stays within a bounded residual of the kinematics supplied by the monocular body model, and the skeleton comes from that model rather than from radar labels, so the fit corrects the video's motion rather than recovering it independently. Elevation is the axis the scaffold reconstructs least well, which is why HuPR's vertical board and the RT-Pose cascade score below HuPR's horizontal board on fidelity. No public simulator takes a capture as input, so RF-Genesis in the comparison is driven forward from the registered motion rather than fitted to the recording being scored.