| Dataset 2 LHI_UIUC_Sport_Activity_10 |
| Contents and Statistics : |
| Category |
Number of image |
Free |
Total |
| Badminton |
10 |
314 |
| Bocce |
10 |
137 |
| Croquet |
10 |
369 |
| Hurdles |
10 |
176 |
| IceSkate |
10 |
310 |
| Polo |
10 |
183 |
| Rockclimbing |
10 |
194 |
| Rowing |
10 |
430 |
| Sailing |
10 |
199 |
| Snowboarding |
10 |
200 |
|
Matlab codes to load this category (see MatlabToolbox page for more information):
clear; close all;
HOMEIMAGES = 'C:\Yourdatabase\LHI_UIUC_Sport_Activity_10\Images';
HOMEANNOTATIONS = 'C:\Yourdatabase\LHI_UIUC_Sport_Activity_10\Annotations';
HOMELABELMAPS = 'C:\Yourdatabase\LHI_UIUC_Sport_Activity_10\LabelMaps';
Code to load the database and mapping the names
rawDB = LHIdatabase(HOMEANNOTATIONS); NEWHOMELABELMAPS = 'C:/temp/yourposition/';D=LHIregulatenames(rawDB,HOMELABELMAPS,NEWHOMELABELMAPS,'namelist_sportactivity.txt'); |
 Figure1 Name table of all categories and correspond color of in labelmaps. Left three columns are R,G,B values respectively.
|
|
Code to produce the figure below:
LHIobjectnames(D);
|
Figure 2 Object frequency count |
|
View one image for hierarchical decomposition
Code to produce the figure:
LHIdbshowimage(D,3,HOMEIMAGES);
 |
Figure 3. Hierarchical decomposition of an image |
|
Display groundplane and/or horizon lines.
Some images have parallel lines on the ground (e.g. the mark on the court in Figure 4). For these images, ground plane can be defined by two pares of parallel lines and corresponding vanishing points (intersections of parallel lines). When image does not explicitly contain geometrical/structural evidence (see Figure 5), which is the more general case, a horizon line imagined by the labeler (human can estimate the position of horizon line pretty well).
Code to produce figure:
LHIdbshowgeometry(D,7,HOMEIMAGES); |
 |
Figure 4. The intersection of two pares of parallel lines (red, blue) are vanishing points (blue dots). The horizon line (green line) is the connection between two vanishing points. |
Code to produce figure:
LHIdbshowgeometry(D,16,HOMEIMAGES); |
|
Figure 5 No strong clue of parallel lines and vanishing point, horizon line (green line) is conjectured |
|
Template of human skeleton
To analysis sport activity categories, it is desirable to know something about the human body gesture. Therefore, this subset is particularly labeled with a skeleton template. See the following Figure 6., a template is a graph with a common set of keypoints. You can use this feature to position any body part you are interested in (e.g. No.17 for right knee, No.23 for left foot), or to find out what is the direction of the torso (link between No.26 and No.27)
Code to produce figure:
badminton = LHIquery(db,'folder','hurdle');
db_part = LHIquery(badminton, 'object.parts.name', 'skeleton');
LHIdbshowparts(db_part, HOMEIMAGES, 1, [2,2]); |
 |
Figure 6. Template of human skeleton. The left figure represent the skeleton template. Each keypoint (green point) is marked with a digit indicating the point id. |
|
Statistics of the category
General statistics including number of keypoints on the boundary for each object, sum of object/image area ratio. Position (x,y axis of center point) histogram.
Code to produce figure:
[objectnames, instancecounts, areacounts, pointcounts, positions]=LHIobjectstats(D, HOMEIMAGES, NEWHOMELABELMAPS); |
 |
Figure 7 Histogram of the number of key points used to define each object |
 |
Figure 8. Histogram of the percentage of pixels occupied (relative to the image size) by each object instance |
 |
Figure 9. This plot shows the distribution of locations occupied by each instance. Each dot corresponds to the original location, relative to the image frame, of each object instance. These plots help to understand some of the biases that the photographers might have when taking pictures of specific objects. |
|
Related Publications:
L.-J. Li and L. Fei-Fei. What, where and who? Classifying event by scene and object recognition. Proc. Int'l Conference on Computer Vision (ICCV), 2007 |
|