Skip to main content
Back

MC-GLM Adds Real-Time Uncertainty to Self-Driving Car Detectors Without Retraining

MC-GLM adds calibrated uncertainty estimates to any trained 3D object detector without retraining, a critical step for safe autonomous driving deployment.

  • MC-GLM adds per-object uncertainty estimates to any pre-trained 3D object detector with no retraining required.
  • The method combines the Laplace approximation (fitting a Gaussian over trained weights) with a linearized model to propagate uncertainty to each bounding box.
  • Key innovation: the number of Monte Carlo samples is constant and independent of how many objects are in a scene, enabling full GPU parallelization.
  • Validated on the nuScenes dataset with the CenterPoint LiDAR detector, showing well-calibrated uncertainty quality metrics (AvU, P(U|I), P(A|C)).
  • Practical use cases include sensor fusion weighting, conservative safety planning, and active learning data selection.
  • No public code released yet; the laplace library is a compatible starting point for implementation.

Every time an autonomous vehicle's perception stack spots a pedestrian or a truck, it outputs a bounding box. What it almost never tells you is how confident it is in that box. A new paper from researchers at introduces MC-GLM (Monte-Carlo Generalized Linearized Model), a method that bolts calibrated uncertainty estimates onto any pre-trained 3D object detector, no retraining required. The target: safety-critical systems that need to know not just what is out there, but how sure the model is about each individual detection.

Why bounding boxes alone are not enough

Object detection is a fundamental task in autonomous driving, and beyond bounding-box predictions, uncertainty estimation is necessary for safe and trustworthy perception. A detector that is wrong but confident is arguably more dangerous than one that is wrong and knows it. Planners and safety monitors downstream need a per-object uncertainty signal to decide whether to brake, yield, or flag for human review.

The challenge is that most production detectors are deterministic: they were trained once, deployed, and never touched again. Deep neural networks are prone to make overconfident predictions on outliers, such as out-of-distribution data and data under distribution shift, which is especially harmful in safety-critical applications. Retrofitting uncertainty into these frozen models is the core problem this paper tackles.

The existing toolkit and its gaps

There are three main families of approaches to uncertainty quantification (UQ) in neural networks, each with a real-world catch:

  • Bayesian Neural Networks (BNNs): Principled but computationally prohibitive for large modern networks. Bayesian Neural Networks provide a reliable approach, but it often becomes computationally prohibitive to apply them to modern large-scale neural networks.
  • Monte-Carlo Dropout / Deep Ensembles: Run the model multiple times with random dropout or train several copies. These are sampling-based and require modifying training, making them not fully post hoc.
  • Linearized inference (GLM): Approximate the network as a linear model around its trained weights using the Laplace approximation. Fast and post hoc, but requires one backward pass