Related Publications
This set of data is used in the following publications:
- IEEE T-PAMI 2007: Rapid Biologically-Inspired Scene Classification
Using Features Shared with Visual Attention
- IROS 2007: Biologically-Inspired Robotics Vision Monte-Carlo Localization in the
Outdoor Environment
- ICRA 2008: Storing and Recalling Information for Vision Localization
- VSS 2008: Comparisons of Gist Models in Rapid Scene Categorization Tasks
Dataset
There are three sites on the USC campus:
- 1. Ahmanson Center for Biological Science (ACB)
- 2. Associate and Founders Park (AnF)
- 3. Frederick D. Fagg park (FDF)
These sites are selected because they present a different challange visually: a human made building complex, a park full of trees, and an open field area. The sites can be viewed in the following map (figure 1).
Figure 1. Map of testing sites
The data is taken using a camera.
The videos are arranged as follows: each column lists videos from a particular path segment (there are 9 for every site, observe the site map for the segment locations). Each row is designated for a particular trial run (which means that a row has the same lighting condition), except for ACB. This is because ACB is the first testing site and, at that time, we were still trying to figure out the best way to train a neural network that would have lighting invariance.
Selection of filming
times that include all lighting conditions is the main issue in collecting training samples. Because lighting space is
hard to gauge, it takes several days to come up with times of day to
cover most of them.
At the end settle on the following times: early morning, noon, early afternoon, late afternoon, and early evening. We do not include the night time videos because some sites are simply too dark.
For testing episodes, we tape two dark and two bright
conditions separately (they are not used for training). Also note that there are also other
significant discrepancies in forms of people walking, missing objects
(a park bench taken away), and new objects (service vehicles) between the training and testing dataset.
Ahmanson Center Of Biological Science (ACB)
The ACB scenes (shown in figure 2) are filmed throughout
the narrow corridors of the building complex. Most of the surroundings
are flat walls with little texture and solid lines that delineate the
walls and different parts of the buildings.
 |
 |
| Figure 2. Map and example images of the ACB site |
Figure 3 shows the different lighting conditions tested, from left to right: late afternoon (trial 1), early evening with
lights already turned on (trial 2), mid-afternoon (trial 3), and
noon (trial 4).
Figure 3. Lighting conditions for the ACB site testing
The entire ACB data in one .tar file: ACB.tar.gz
Associate and Founder's Park (AnF)
The scenes of AnF (figure 4), are dominated by vegetations.
There are fewer extractable structures as large areas of the images
are practically un-segmentable as leaves overrun most regions. In
addition, the length of the segments at this site are about twice the
lengths of the segments in ACB.
 |
 |
| Figure 4. Map and example images of the AnF site |
Figure 5 shows the different lighting conditions tested, from left to right: overcast (trial 1), early evening with lights
already turned on (trial 2), mid-afternoon (trial 3), and noon
(trial 4).
Figure 5. Lighting conditions for the AnF site testing
The entire AnFpark data in one .tar file: AnFpark.tar.gz
Frederick D. Fagg (FDF)
A large portion of the FDF site (figure 6) is sky, mostly textureless space with scattered light clouds. The main reason for testing at this site is to assess the gist response on sparser
data. The length of the segments are about 50 percent longer than the ones in the second experiment (three times that of ACB).
 |
 |
| Figure 6. Map and example images of the FDF site |
Figure 7 shows the different lighting conditions tested,from left to right: early evening (street lights are off), early evening (street lights are already on), noon,
and middle of afternoon, with the dramatic long shadow cast on the field by a
tall building.
Figure 7. Lighting conditions for the FDF site testing
The entire FDFpark data in one .tar file: FDFpark.tar.gz
Copyright © 2000 by the University of
Southern California, iLab and Prof. Laurent Itti