RainSense
Last updated: 2026-05-09
RainSense is a multi-sensor autonomous-driving perception dataset collected under natural rainfall with fine-grained rainfall intensity annotations. Its main value is not dense point-wise rain labels, but calibrated rain-rate stratification across camera, LiDAR, and 4D radar data.
Related pages: dataset index, LiDAR artifact removal techniques, radar-LiDAR fusion in adverse weather, production perception systems
What It Measures
RainSense measures how perception sensors degrade as natural rainfall intensity increases. The SAE record describes 728 ten-second cases across five conditions:
| Condition | Cases |
|---|---|
| Clear | 145 |
| Light rain | 214 |
| Moderate rain | 204 |
| Heavy rain | 98 |
| Torrential rain | 67 |
The dataset uses a laser-optical disdrometer to measure rainfall intensity in mm/h while synchronized sensor data is recorded. This makes RainSense useful for setting quantitative operating thresholds such as "LiDAR confidence falls below target at this rain intensity, but radar remains stable."
Sensors And Modalities
| Modality | Notes |
|---|---|
| Camera | Image degradation and blur under rainfall |
| LiDAR | Point sparsity and weakened returns under rain |
| 4D mmWave radar | Weather-resilient comparison channel |
| Laser-optical disdrometer | High-precision rain intensity measurement, identified as Parsivel-2 in the GitHub README |
| Static/dynamic targets | Dummy target placed at different distances in representative scenes |
Labels And Tasks
| Label type | Use |
|---|---|
| 2D target boxes | Camera detection degradation by rain level |
| 3D target boxes | LiDAR/radar target detection under rain |
| Rain intensity annotation | Response curves by mm/h and rain class |
| 10-second case windows | Sensor stability and temporal degradation |
RainSense supports sensor performance characterization, SOTIF trigger analysis, target detection under rain, and radar-vs-LiDAR robustness comparisons.
Weather And Environment
The SAE abstract states that RainSense was recorded in nine representative campus intersection scenarios under natural rain. The GitHub README describes five rainfall levels: clear, light, medium, heavy, and torrential.
Because rainfall is natural and measured, RainSense is more suitable than synthetic data for validating rain-rate thresholds. It is less suitable for broad object taxonomy testing because the SAE abstract describes a single dummy target placed at various distances.
Benchmark Use For Perception And Removal
Use RainSense to validate:
- camera blur and detection loss as rainfall increases;
- LiDAR point-count and intensity degradation by rain class;
- radar stability across rain levels;
- rain-aware confidence calibration;
- operational rain thresholds for slow, stop, or radar-primary modes.
For removal validation, RainSense should be paired with REHEARSE-3D. Use REHEARSE-3D for point-wise rain removal, then use RainSense to check whether the cleaned perception stack behaves better under real rainfall intensity bins.
Strengths
- Natural rainfall rather than only simulated rain.
- Fine-grained rain intensity measurement with a disdrometer.
- Camera, LiDAR, and 4D radar recorded synchronously.
- Clear distribution across rain levels.
- Open GitHub repository and release metadata are available.
Limitations
- The SAE abstract describes a single dummy target rather than dense real traffic.
- It is intersection/campus data, not full-route driving.
- It is rain-focused and does not cover snow, fog, dust, steam, or de-icing spray directly.
- Labels are target boxes, not dense point-wise rain/no-rain annotations.
- Published paper access is through SAE; verify license and download terms before redistribution.
Airside Transfer
RainSense is useful for airside rainfall ODD definition:
- define rain-rate thresholds for LiDAR degradation;
- validate radar as the stable modality under heavy/torrential rain;
- test target detection at known ranges under measured rainfall;
- calibrate perception health monitors against mm/h rather than vague weather tags.
It is a weak proxy for de-icing mist because the particle type, temperature, and sensor-window contamination differ from natural rain. Treat it as a rain-intensity baseline, not a glycol-spray validation set.