WarNav: An Autonomous Driving Benchmark for Segmentation of Navigable Zones in War Scenes

TL;DR

WarNav introduces a real-world, war-zone semantic segmentation benchmark aimed at robust navigability for autonomous ground vehicles under data-scarce, ethically constrained conditions. Built from the DATTALION repository, WarNav emphasizes frugality in annotation by annotating a small subset of validation/test images while keeping the training set unannotated, and it defines a tailored wmIoU metric that prioritizes near-field navigable regions. The authors benchmark multiple models trained on urban datasets, analyze cross-domain generalization, and demonstrate that fusing Cityscapes, RUGD, and Earthquake data yields the best WarNav performance, highlighting the value of diverse outdoor domain knowledge for destruction-era environments. The work provides open data and methodological guidance for future research in extreme deployment contexts, and points to unsupervised domain adaptation as a promising avenue to further close the domain gap with minimal annotation. Overall, WarNav advances safe autonomous navigation in conflict-affected environments by offering a standardized, frugal evaluation framework and actionable insights for robust perception under severe data constraints.

Abstract

We introduce WarNav, a novel real-world dataset constructed from images of the open-source DATTALION repository, specifically tailored to enable the development and benchmarking of semantic segmentation models for autonomous ground vehicle navigation in unstructured, conflict-affected environments. This dataset addresses a critical gap between conventional urban driving resources and the unique operational scenarios encountered by unmanned systems in hazardous and damaged war-zones. We detail the methodological challenges encountered, ranging from data heterogeneity to ethical considerations, providing guidance for future efforts that target extreme operational contexts. To establish performance references, we report baseline results on WarNav using several state-of-the-art semantic segmentation models trained on structured urban scenes. We further analyse the impact of training data environments and propose a first step towards effective navigability in challenging environments with the constraint of having no annotation of the targeted images. Our goal is to foster impactful research that enhances the robustness and safety of autonomous vehicles in high-risk scenarios while being frugal in annotated data.

Figures (3)

• Figure 1: Histogram of number ($\times 10^7$) of pixels per class for the test and the validation sets of WarNav. When ignoring 'overlay', the 5 remaining classes constitute the so-called $L_5$ setting used in this paper.
• Figure 2: Histogram of connected regions per class for the test and the validation sets of WarNav.
• Figure 3: Illustration of the influence of the training datasets. Columns from left to right are: test images of WarNav, their corresponding annotations, predictions of SegFormer(MIT-B5) trained on Citysapes, RUGD, Earthquake and the combination of the three datasets.