Lightweight Uncertainty Quantification with Simplex Semantic Segmentation for Terrain Traversability

TL;DR

The paper addresses the need for fast, reliable per-pixel uncertainty in terrain-traversability segmentation for robotics. It introduces a lightweight, architecture-agnostic Simplex Semantic Segmentation module that attaches to pretrained segmentation models and uses $N$ class prototypes arranged on the $(N-1)$-dimensional unit hypersphere; per-pixel uncertainty is the distance to the nearest prototype, learned via a cosine loss that aligns feature projections with class prototypes. Empirical results on the off-road Rellis3D dataset and multiple external datasets show improved out-of-distribution uncertainty over a DeepLabV3+ baseline, with informative per-pixel visualizations and favorable inference costs compared to MC dropout. The approach offers a practical, plug-in solution for uncertainty-aware, real-time navigational decisions in robotic systems.

Abstract

For navigation of robots, image segmentation is an important component to determining a terrain's traversability. For safe and efficient navigation, it is key to assess the uncertainty of the predicted segments. Current uncertainty estimation methods are limited to a specific choice of model architecture, are costly in terms of training time, require large memory for inference (ensembles), or involve complex model architectures (energy-based, hyperbolic, masking). In this paper, we propose a simple, light-weight module that can be connected to any pretrained image segmentation model, regardless of its architecture, with marginal additional computation cost because it reuses the model's backbone. Our module is based on maximum separation of the segmentation classes by respective prototype vectors. This optimizes the probability that out-of-distribution segments are projected in between the prototype vectors. The uncertainty value in the classification label is obtained from the distance to the nearest prototype. We demonstrate the effectiveness of our module for terrain segmentation.

Figures (6)

• Figure 1: This figure describes our ability to predict segmentation classes with uncertainty. The input image consists of regions that are in-distribution and areas that are out-of-distribution. We are able to extend a traditional segmentation model (a) by estimating uncertainty through distances to the nearest prototype vector (b). In this example, we are able to detect out-of-distribution areas (fire) through our uncertainty estimation.
• Figure 2: Segments from SceneParse150 for which our model (trained on Rellis3D) assigns the highest uncertainties.
• Figure 3: Our model (middle) yields high uncertainty values for the tree, pigs and fence, whereas the uncertainty obtained with the standard method is scattered (right).
• Figure 4: In-domain cases from the Rellis3D dataset. The segmentation predictions are often correct (third column) and larger uncertainty values are assigned to unclear boundaries between segments for the standard method (last column) than our method (fourth column).
• Figure 5: Visualisation of the uncertainty for images with out-of-domain regions: two images from our dataset (fog and fire) and four images from the SceneParse150 sceneparse150 dataset (building, house and countryside).