This repository constitutes the WarNav dataset, made of annotations of selected subsets of images provided by the open-source DATTALION multimedia repository. WarNav is specifically tailored to enable the development and benchmarking of semantic segmentation models for autonomous ground vehicles in unstructured, conflict-affected environments.
The following is provided in Google Drive:
- 3 files providing the names of pictures selected from Dattalion. One file is given per dataset split (training, validation and test sets).
- train_dataset_selection.txt (5354 image names)
- validation_dataset_selection.txt (100 image names)
- test_dataset_selection.txt (100 image names)
- Validation and test annotation, in Cityscape format
- Validation_annotation
- Test_annotation
Among the 100 validation dataset images, 3 images do not contain any pixel of the classes of interest. Thus, annotation of validation set provides information for the 97 images which have at least 1 pixel of a class of interest.
The Test_annotation provides a single annotation for 90 images, and redundant annotations given by 3 different annotators for 10 additional images. For performance evaluation, the annotations of Annotator 2 are used.
Annotated classes of interest are the following:
- Overlay: Regions containing graphical overlays or annotations that were added post-capture. These pixels are excluded from both training and performance evaluation, as they do not correspond to real-world scene contents.
- Road: Surfaces intended for civilian vehicular traffic, typically paved with asphalt or similar materials.
- Drivable: Areas that are not formal roads but are deemed traversable by military 4x4 vehicles (e.g., dirt paths, open fields).
- Pedestrian: Humans. Accurate detection of this class is essential for tasks related to safe autonomous navigation.
- Vehicle: Civilian vehicles that are potentially operable. Obstacle avoidance algorithms would consider them as potentially non static obstacles. Damaged or abandoned car wrecks are excluded from this category.
- Background: All non-annotated regions are classified as background, encompassing areas where navigation is not feasible (e.g., buildings, vegetation, sky, rubble, or other obstacles).
Data annotations are under Creative Commons Attribution Non Commercial 4.0 license.
Associated paper can be read here (to be filled later).