Invascal makes LiDAR segmentation uncertainty-aware without performance loss
A new method lets self-driving cars know when they're unsure—without slowing down.
Researchers from the automotive and robotics domain introduce Invascal, a self-calibration technique that makes LiDAR semantic segmentation uncertainty-aware without the computational cost of Monte Carlo dropout or ensembles. The key is a lightweight Adapter Head that splits the prediction into two streams: a Preference Head for class ranking and a Strength Head for refining uncertainty. This enables evidential Dirichlet representations that naturally model prediction confidence. The inverse-vacuity objective directly supervises the strength signal to prevent overconfident predictions and runaway evidence growth, delivering well-calibrated uncertainty estimates in real time.
Evaluated across multiple datasets and backbone architectures, Invascal consistently outperforms deterministic baselines, MC dropout, and prior evidential methods in calibration metrics while preserving competitive segmentation accuracy—where earlier evidential approaches often degrade performance. The method adds minimal inference overhead, making it suitable for autonomous vehicles and mobile robots that must operate safely under uncertainty. Accepted at IEEE ITSC 2026, the work offers a practical plug-and-play solution for any LiDAR segmentation pipeline.
- Invented an architecture-agnostic Adapter Head with Preference and Strength branches for evidential Dirichlet representations.
- Inverse-vacuity self-calibration objective prevents overconfidence and runaway evidence growth without extra compute.
- Matches or beats MC dropout and ensembles on calibration, while retaining accuracy that prior evidential methods lose.
Why It Matters
Self-driving cars and robots can finally know when their perception is unreliable—key for safe autonomous operation.