DyRAD: Radar Novel View Synthesis for Dynamic Driving Scenes

DyRAD. Radar Novel View Synthesis for Dynamic Driving Scenes. By Merav Keidar and three others, from the Technion, Cornell Tech, and one other institution. Published on the thirtieth of September. Its code repository is worked out: about four hundred source files, with a readme and a licence.

Rebuilding a dynamic driving scene from recorded sensor data lets autonomous driving systems be evaluated in closed loop, because the reconstruction can synthesize what the sensors would have seen off the original trajectory. Radar has something cameras and LiDAR lack: it measures radial velocity directly, through Doppler. Existing radar novel-view synthesis does not make use of that. Methods that handle dynamic scenes reconstruct only range and azimuth, while methods that do render Doppler assume the scene is static. There is a second flaw. Radar processing spreads every reflection across several bins, and existing representations soak that spread into the scene geometry itself, so the scene renders wrongly once the viewpoint moves. What was really a property of the sensor gets stored as if it were a property of the world, and it is carried into new views where it no longer belongs.

DyRAD addresses both. It models a driving scene as static background reflectors plus dynamic point reflectors that are tracked as they move, and renders complete range-azimuth-Doppler tensors. The key separation is between the scene and the sensor. Reflectors are rendered through a fixed, analytic point-spread function derived from the radar's own signal-processing chain, so the spread the sensor introduces is never baked into the representation of the scene.

Motion enters through velocity. Each reflector's velocity comes from object tracks and is projected onto the line of sight, which makes Doppler two things at once: something the model renders, and a source of supervision for those very tracks. The separation of sensor from scene pays off beyond reconstruction quality. Because the scene holds no sensor-specific spread, the same reconstruction can be rendered under different radar specifications without being refitted, which gives zero-shot transfer across sensor configurations.

DyRAD is evaluated on RADIal, on Boreas, and on a synthetic benchmark. The tests include poses on the recorded path and also displaced viewpoints, which prior work had left untested. On RADIal, DyRAD recovers radar detections for ninety point seven percent of the objects detected in the reference. The strongest baseline recovers twenty-six point nine percent. Both figures count the same reference objects, so the comparison reads directly as how many real targets survive into the rendered radar view: the large majority for DyRAD, about a quarter for the best of the rest.

Two choices carry this design. Treating Doppler as both output and supervision puts radar's distinctive measurement to work instead of discarding it. Keeping the point-spread function analytic and outside the scene stops a property of the sensor from masquerading as geometry, and that is what lets one reconstruction serve many radars.