Experiments on GT Data

Before attempting to decipher this page please review the data description page, our method page, and peruse the FTP data site.

Available Data

As noted in the data description, there are 15 subjects for whom we have all forms of video data: angle-indoors, side-indoor-near, side-indoor-far, and angle-outdoor.  For each subject and for each of the two angle conditions, we have 6 trials each.  For the side conditions, we have 3 trials each.  This means that there is a total of: 15*(6+6+3+3) = 270 trials.  (Actually, there are 268, see below.)

On the ftp data site, all data is organized by subject.  In each subject directory is the outdoor angle directory and the indoor directory which has both side and angle data. 

Now the bad news.  The angle for the indoor data is at ground level with the subject walking at  45 degrees with respect to the camera.  Because of that symmetry, we didn't distinguish between the person walking left to right or right to left, or (even worse) toward the camera or away.  Also, though there are six side views per subject (3 close, 3 far) the six side view files in each directory are not in any particular order.  Furthermore, some are right to left, some are left to right.  

Data Anomalies

• Subjects 9 and 10 have only 5 indoor angles sequences

Trial Labeling and Data Files

In light of this data structuring, we generate a table that is a similarity comparison (by finding the mahalanobis distance) between the indoor angle trials and all trials (including the indoor angle). The covariance matrix used to compute the mahalanobis distance is the individual variation of the feature (walk) vector. To compute this matrix , we subtract the mean walk vector of each subject from each of their trials and then compute the covariance over all the trials.

The table is 268 rows by 88 columns. The headings of the columns refer to the gallery (angle-indoors), and the row headings are the probe views (angle-indoors, side-indoor-near, side-indoor-far, and angle-outdoor). For out technique there is not a training set.

The headings for the rows and columns will be written using a TRIAL label defined as follows:

<TRIAL> := sub<SUB#><COND><TRIAL#>

<SUB#> := 02 | 03 | 04 | 05 | 06 | 07 | 08 | 09 | 10 | 12 | 15 | 16 | 17 | 18 | 19

<COND>:= ai | si | ao

<TRIAL#>:= t01 | t02 | t03 | t04 | t05 | t06

where,

• ai: angle-indoors
• si: side-indoor
• ao: angle-outdoor

 Results

1. As stated above, we generated a table that is a similarity comparison between the indoor angle trials (gallery) and all trials (including the indoor angle). The similarity matrix is here as an Excel spreadsheet, and text file with just numbers:
   • Walk vector w [xls, txt].
     • cmc-curve of indoor angle vs. outdoor angle
     • cmc-curve of indoor angle vs. indoor side far
     • cmc-curve of indoor angle vs. indoor side near
   • Walk vector s [xls, txt].
   • Mappings between row and column headings and filenames
     • columns (Gallery)
     • rows (probe)

2. Also, we generated a table that is a similarity comparison between the indoor-angle (gallery) mean trial per person and all trials (including the indoor angle). The similarity matrix (15 columns and 268 rows) is here as an Excel spreadsheet, and text file with just numbers:
   • Walk vector w [xls, txt].
   • Walk vector s [xls, txt].
   • Listing of column and row headers
     • columns (Gallery)
     • rows (probe)

Experiment Specifics

• Subject 03 was used to calibrate the depth compensation (scaling) for all camera set-ups.  Subject 03's height is 175.25cm
• All subjects were used to compute "cross-condition mapping functions".  That is, a single linear regression was done for each measured feature in our method that would map from one condition to the other.  All subjects were used because of the limited number of subjects.  

MAIN PAGE