WIT-UAS (Wildland-fire Infrared Thermal UAS Dataset)
CMU AirLab · 2023 · datasets.bot · datasets.bot page
One-liner. High-altitude aerial long-wave infrared (LWIR) thermal dataset from CMU AirLab for detecting crew and vehicle assets in prescribed wildland-fire scenes, with hand-labeled thermal images plus full ROS bags from UAV flights over fire.
Setup
- Datasets / benchmarks: WIT-UAS is a wildland-fire long-wave infrared (LWIR) thermal dataset collected by CMU's AirLab (Airborne Robotics Lab) to detect crew and vehicle assets from aerial views amidst prescribed burns. It is split into two subsets: WIT-UAS-ROS (full ROS bag files containing sensor and robot data of UAS flights over fire) and WIT-UAS-Image (hand-labeled LWIR images extracted from the bags at ~1 frame per second). The dataset contains 6,951 total thermal images, of which 2,062 are manually bounding-box annotated, yielding 5,030 labeled vehicles and 1,542 labeled humans. The dataset.classes file defines the categories: person, car, bicycle, othervehicle, plus noobject and dontcare. Data was collected over three prescribed fire seasons (fall 2021, spring 2022, fall 2022) at Pennsylvania State Game Lands (SGL 174 near Rossiter, SGL 111 near Confluence, SGL 42 in Reade Township, all in Western PA) using DJI M100 and DJI M600 UAVs, with thermal sensors connected to an onboard NVIDIA Jetson Xavier NX. The dataset addresses the problem that thermal detectors trained without fire data frequently misclassify flames as people; it is the first public LWIR dataset focused on assets near fire. Code, pretrained YOLO/SSD models, and download scripts (via minio) are released on GitHub under GPL-3.0. Published at IEEE/RSJ IROS 2023. License: custom. Download: https://github.com/castacks/WIT-UAS-Dataset.
- Hardware / simulator: Embodiment: drone. Environment: outdoor. Realness: physical.
Schema
Two subsets: WIT-UAS-ROS (full ROS bag files of UAS flight sensor/robot data) and WIT-UAS-Image (6,951 LWIR thermal images, 2,062 hand-labeled with bounding boxes). Classes: person, car, bicycle, othervehicle (plus noobject, dontcare). 5,030 vehicle labels, 1,542 human labels. Images sampled at 1 fps. Download via python scripts/download_data.py (minio).
Links