Deep learning models for 3D perception often fail when moved from one sensor to another. A model trained on a 64-beam Velodyne sensor might struggle to recognize a simple pedestrian when deployed on a 32-beam Ouster. While much research focuses on geometric differences like point density, a more insidious problem lies in LiDAR intensity—the strength of the laser pulse return.
The Problem with Raw Intensity
LiDAR intensity is not a universal physical constant. As noted by Gerhard Kurz and Peter Biber in When Geometry is not Enough: Using Reflector Markers in Lidar SLAM, intensity values lack standardization across manufacturers. A value of "150" on one sensor might represent a highly reflective traffic sign, while on another, it might represent a dark asphalt road.
This inconsistency stems from three primary factors:
- Hardware Calibration: Manufacturers use proprietary scales and internal optics.
- Distance Decay: Signal strength weakens as the distance to the object increases.
- Incidence Angle: A laser hitting a surface at a shallow angle returns less energy than one hitting it head-on.
The "Back-to-Basics" Solution: Feature Exclusion
Because intensity is so sensor-specific, the most effective strategy for Domain Generalization (DG) is often to simply delete it. Researchers like Konrad Schindler and Xiao Xiang Zhu found that "back-to-basics" strategies—focusing purely on stable geometric coordinates—outperform complex adaptation algorithms.
In Domain generalization of 3D semantic segmentation in autonomous driving, Jules Sanchez and colleagues demonstrate that excluding reflectivity forces models to learn structural features that are invariant to hardware changes. This prevents the model from overfitting to the "intensity signature" of a specific training sensor.
Moving Toward Calibrated Reflectivity
While removing intensity improves generalization, it also discards valuable data. Intensity is crucial for identifying low-reflectivity objects like pedestrians or lane markings. To bridge this gap, recent work focuses on transforming raw intensity into calibrated reflectivity.
In Reflectivity Is All You Need!, Kasi Viswanath and Srikanth Saripalli propose accounting for range and incidence angles to recover the true material properties of a surface. Similarly, the DIG method introduced by Jinhao Deng in the Cross Mechanism Dataset (CMD) explicitly normalizes intensity alongside density and geometry.
By shifting from raw, sensor-dependent signals to normalized physical properties, researchers are enabling 3D models to maintain high performance across diverse hardware platforms and environmental conditions.