The most complete source of information is the attached file. The original
of that file is in the project library

/projectlibrary/900260.AVED/Project.Documents/Project.Data/Classification
Results/benthic recognition results march 2005.xls

Basically, the procedure was as follows:

A tape from ventana dive 2526 was "digitized" via our betacam video deck and
capture software.

Segments of about 1 min were analyzed with AVED. The files have the name
<dive number><tape number><time code of the first frame>

8 segments were analyzed. This same tape, and these portions of the tape,
had been professionally analyzed in the video lab, so we were just comparing
AVED events with what we already knew from the lab. The animals we were
analyzing were easy to see and identify, so we did not really need that
professional annotation, but we had it anyway.

For each AVED event in the segments shown in the attached .xls file, we
identified the event per the professional annotation: Leukothele,
rathbunaster, or something else ("other"). For each event, we looked at each
individual frame (events have lots of frames, in general), and we "voted"
the frames -- each frame was classified by the classification program, and
we noted the classification for each frame. We tallied the classification of
each frame, and classified the event by "majority rule". All the details of
that "voting" is in the spreadsheet.

We compared the program classification with the professional. We called the
classification "right" if the program and professional agreed. We called the
classification "wrong" if they disagreed. Those were marked with a "1" in
the appropriate column for each event.

We tallied all the classifications, to measure percentages, total instances,
total images.

In Ishbel's notebook we have the final calculations of Precision and Recall,
and the confusion matrix, for these 210 events. The analysis is called
"avedClassifierStats" and it probably on the project library data page
someplace. I can show you the sheet in any case.


I think that is all you need now. We have all these images on line and on
data DVD as original single images (in color), and after converting to
grey-scale.