Overview
Algorithms designed for time-evolving domains have a history of being evaluated on synthetic data, e.g., the “moving hyperplane” class of artificial data, where concept drift is introduced manually and any correlation to real-world problems is unestablished. This motivated the creation of natural datasets taken from the problem domain. The natural datasets used here are taken from logged field tests conducted by DARPA evaluators, and have been shown to contain time-varying (drifting) concepts.
Representative Images
Scenarios and Lighting Conditions
Overall, three scenarios are considered. Each scenario is associated with two distinct image sequences, each representing a different lighting condition. There are thus six datasets total. The terrain appearing in the datasets varies greatly, and includes various combinations of ground type (mulch, dirt); foliage; natural obstacles (trees, dense shrubs); and man-made obstacles (hay bales). Lighting conditions range from overcast with good color definition (e.g., DS1B, shown above), to very sunny, causing shadows and saturation (e.g., DS2A). Additional descriptions and representative images from each dataset are available in the related article.
Hand-Labeling
Each dataset consists of a 100-frame hand-labeled image sequence. Each image was manually labeled, with each pixel being placed into one of three classes: Obstacle, Groundplane, or Unknown. If it was difficult for a human to tell what a certain area of an image was—even when using higher-level context—then that region was labeled as Unknown. On average, approximately 80% of each image was labeled as either Obstacle or Groundplane, with the remaining 20% labeled as Unknown.
Working with the Datasets
These are MATLAB-6 compatible .mat files, read using the load() function. Each MAT file represents one frame from the robot log files and contains the raw RGB image, disparity information, and an integer mask with a pixelwise labeling. The disparity information is available for your own stereo processing.
0: ground plane1: obstacle2: this pixel was not labeled by a human
Unlabeled areas have meaning: they may be regions for which the terrain class was hard to tell, even with context, or they may be “don't cares,” such as sky.
More Information
For further information on these datasets, including additional representative images, see:
Michael J. Procopio, Jane M. Mulligan, and Greg Grudic. Learning Terrain Segmentation with Classifier Ensembles for Autonomous Robot Navigation in Unstructured Environments. Journal of Field Robotics (2009). PDF.
Acknowledgements
Special thanks to Wei Xu (at the University of Colorado at Boulder) and to Sharon Procopio for their assistance in labeling these images.
Citing
If you use this data in your research, we ask that you cite it as follows:
@misc{Procopio-LabeledLAGRData-07,
author = {Michael J. Procopio},
title = {Hand-Labeled {DARPA} {LAGR} Datasets},
howpublished = {Available at \url{https://mikeprocopio.com/labeledlagrdata.html}},
year = {2007}
}Download
The datasets can be downloaded as individual ZIP archives using the links below. File sizes are shown in MiB or GiB. Download SHA-256 checksums to verify the archives.
Additional Unlabeled Data
Additional unlabeled frames are available for DS1A, DS1B, and DS2A. These frames occur later in the respective robot test runs and exhibit stronger degrees of concept drift, including terrain not present in the earlier frames (1 to 100). These supplemental datasets start at frame 101, picking up where the labeled datasets above leave off.