🔬 RESEARCH
- RadarScenes study finds point density drives radar-only classification accuracy
A RadarScenes study found point density, not MLP architecture, drove radar-only classification accuracy.- The findings target autonomous-driving perception and radar ML developers.
- A radar-only 5-class classifier used RadarScenes sensor 2, a 3-layer MLP, and 65-dimensional histogram features.
- Macro F1 rose from 0.381 with 1 point per instance to 0.764 with 5 points.
- Fourteen architecture and feature variants stayed within the measured noise floor at fixed point counts.
- The report identifies additional points from multi-scan or multi-sensor data as the direct way to address the accuracy ceiling.