TRECVID 2013 Collaborative Annotation Short Help
Organized by LIG
(Laboratoire d'Informatique de Grenoble)
and LIF
(Laboratoire d'Informatique Fondamentale de Marseille)
More info on
TRECVID 2013 Collaborative Annotation.
Authors: Georges Quénot (LIG) and Stéphane Ayache (LIF)
The TRECVID 2013 Collaborative Annotation system has been designed to be
efficient and easy to use. It operates through the Web and you can access to
it from your favorite browser at:
http://mrim.imag.fr/tvca2013/al.html.
Participation is restricted to groups that are registered TRECVID participants
and that have signed the IACC.1 license agreement
The system operates in a parallel mode in which the images are displayed
by groups in a two-dimensionnal array.
Access page
The access page form has eight inputs:
- Groupname: the name of the group you belong to. It will be given to you by
the contact person for your group. Please refer to the
group table to
identify the contact person for your group.
- Username: an identifier for the person doing the annotations. There should be
one and only one for each physical person doing annotations in your group. It can be
chosen arbitrarily (e.g. steve or user#). It is local to the group and will not be
confused with a possibly identical username in another group. Check with the contact
person for your group to ensure unicity within your group. There can be as many username
as you want in one group. The username is actually used only to ensure a minimal
continuity of the feature to be annotated within an annotation session and to make
statistics about the annotation effort.
- Password: the password for the group you belong to. It is related to the
groupname and not to the username. It will be given to you by the contact person for
your group.
- Keyframes path: This field should contain the path of a local copy of the
set of keyframes (e.g. Z:\tv2013\IACC.1 or /data/tv2013/IACC.1).
If the path is valid and if your browser is properly configured (see below), a local
copy of the keyframes will be used for display. We no longer support the version
with online keyframe download since it is often too slow and unreliable. If the
keyframes does not display properly, you may have specified a wrong path information
and/or your browser is not properly configured.
- Columns and Rows: These entries determine the number of rows and
of columns of images that are displayed for annotation. This allows you to adapt the
number of displayed images to your screen resolution.
- Width and Height: These entries determine the size in pixels of images
that are displayed for annotation. The default size is 224x126 pixels. You may choose
larger values if you have a high resolution screen. You have to adapt both the number
of images and the size of the images to the size of your screen. It may be better to
have fewer but larger images. It is not so efficient to have too many images to annotate
at once.
Using a local copy of the keyframes
The "keyframes" can be downloaded from the LIG site as a
TGZ or
ZIP archive (about 8.5GB).
The "Keyframes path" field should
not include the "keyframes" part (in the above example, the keyframes are in
"Z:\tv2013\IACC.1\keyframes" or "/data/tv2013/IACC.1/keyframes").
The keyframes are distributed as a unix gzip-compressed tar file ("keyframes.tgz")
or a winzip archive ("keyframes.zip"). You need to expand the archive so that the
system can access its contents.
For (apparently good) security reasons, most recent versions of browsers do
not accept to execute/display something with a "file://..."
URL included in something with a "http//..."
URL. You have to
configure your browser so that it allows that, at least for pages/programs
from "http//mrim.imag.fr/"
(the collaboratiove annotation server).
Here is how to proceed:
If you are using a recent version of Firefox (based on Mozilla 1.8 or higher):
(this will probably work for any Mozilla-based Web browser)
- You have to locate your profile directory. It is normally something like:
D:\Documents and Settings\<username>\Application Data\Mozilla\Firefox\Profiles\xxxxxxxx.default
(XP) or
D:\Users\<username>\AppData\Roaming\Mozilla\Firefox\Profiles\xxxxxxxx.default
(Vista)
where xxxxxxxx
is a randomly generated 8-character string.
- You have to put in this directory the following
user.js
file. If you already have there
a user.js
file, you should rather append to it the content of the
user.js
file and if the already present
file already have a similar section, you should simply edit it and add
"http://mrim.imag.fr"
to the list of authorized hosts. DO NOT execute
user.js
in your browser, just save it and
copy it where appropriate or open it with a text editor to copy and paste its
content where appropriate.
If you are using a less recent version of Firefox:
(this will probably work for any Mozilla-based Web browser)
- Type
about:config
in the URL input
- Double click on the line
security.checkloaduri
to set it to false
We strongly recommend however that you ugrade your version of Firefox instead
since this operation will allow any host (not just mrim.imag.fr
) to
execute/display something with a "file://..."
URL included in something with a "http//..."
URL which is
a potential security problem.
If you are using Internet Explorer:
- Click on "Internet Options" in the "Tools" menu.
- Click on "Security".
- Click on "Trusted Sites".
- Click on "Sites".
- Type
"http//mrim.imag.fr"
into the textbox.
- Click the Add button.
- Uncheck the "Require server verification (https://)" box.
- Click on "Close".
- Click on "OK".
Annotation page
Once you click on "Login", you see the typical annotation page. You should choose
"Rows" ans "Columns" values in the annotation page so that you display a maximum
number of images on your screen without having to scroll. You can try different
values and use the "Back" button of your browser to go back to the access page
until you find the optimal values. Nothing will be stored if you do not click
on the "VALIDATE" button.
An array of reduced size images with the chosen number of rows and columns
is displayed on the right part of the screen. Each image comes with a set of
three radio buttons labelled 'N', 'S' and 'P'. Selecting one of them corresponds
to annotate the image as "Negative", "Skipped" or "Positive" for the target
feature. By default, all images are selected as Negative as this is the most
frequent case. There is also a right arrow on the right below each image.
Clicking on it will play the associated shot in a separate window (not
recommended if you have a slow Internet access).
You will need a video player plugin able to display MPEG-4 files for TRECVID 2013.
In order to ease the annotation, the image track of short shots has been slowed down
by frame repilcation while the sound has not been slowed down but extended outside
of the shot.
There are two ways of changing the annotation of an image. The first
one is to check the corresponding button just below the image. The second one
is to click on the image. This will change its annotation from "Negative" to
"Positive", from "Positive" to "Skipped" or from "Skipped" to "Negative".
This is because it is easier and faster to click on the image than on a
radio button and because the first most frequent change is from "Negative"
to "Positive".
When an image is annotated as "Negative", "Skipped" or "Positive",
the background color of the corresponding cell is set to red, yellow or
grenn respectively. This currently works fine with the Firefox browser but
not with Internet Explorer (see the "Bugs" section).
An enlarged (full size) view of any reduced size image can be displayed
in the upper left part of the screen simply by putting the mouse pointer
over the reduced size image.
The feature to be annotated is displayed in the lower left part of the
screen. All the displayed images have to be annotated for this feature.
There are two action buttons in the annotation page:
- VALIDATE: Store the selected annotation for all the currently
displayed images and go to the next series of images. It is very important
to check and properly annotate all the displayed images before clicking
on this button.If anything goes wrong (e.g. images are not displayed),
please DO NOT hit this button and go back to the Login page.
- EXIT: Terminates your current annotation session. The annotations
of the currently displayed images are NOT stored. Please do not use the "Back"
button of your browser afterwards to start new annotations but go through the
http://mrim.imag.fr/tvca2013/al.html
entry page again.
Warning page
Each time the system switches the feature to be annotated (and in the
beginning), a warning message is displayed with two action buttons:
- OK: Acknowledges the warning message and continue.
- EXIT: Terminates your current annotation session. Please do
not use the "Back" button of your browser afterwards to start new
annotations but go through the
http://mrim.imag.fr/tvca2013/al.html
entry page again.
The warning message page was added because feature switches are quite frequent
due to the active learning approach and they are easy to miss otherwise.
General
We are really interested into having a very high quality annotation since we know
that both false positives and false negatives really harm supervised learning
processes. Therefore, please take your time and, if you have any doubt about the
annotation (if you are not sure of what you see or not sure that you fully understood
the limits of the concept definition or ...), please prefer the "SKIP" button.
Please note that, in case of doubt, you can play the shot
associated to a keyframe. It is almost always neccessary in
some cases (like confimation of a positive "Singing"). The rule
is that the shot is considered positive even if the feature
is visible in the shot but not in the keyframe (as it is
judged in TRECVID).
There is an Active Learning engine working in the background of the annotation
system. Its goal is to select first for annotation the samples (keyframes) that
are expected to be the most informative for systems' training. This is normally
transparent to the users. The only expected effect is that we expect to see more
positives in the early stages of the annotation (the effect is not always
terrific...). This process is more efficient if annotations are done in small
chunks. We therefore encourage several small sessions (typically 30-60 minutes
or even shorter) separated by a few hours or one day rather than a continuous
day of work. This is probably better for both the quality of the annotation and
the comfort of the annotators. The minimum volume of annotations to be done by a
single site should not exceed two or three full days of work but doing more is also
encouraged.
Bugs
It sometimes happens that the server fails. The next page may not come or it
may come with no image. If this happens, and only in this case,
please use the "Back" button of your
browser to go the previous page ant try again. We have seen this only when we
use the dynamic loading of images. We recommend the use of a local copy of
the keyframes when possible. This would also speed up the refresh.
It may happen that the system asks you again your groupname, username and
password. In that case, just enter them again an then continue.
When an image is annotated as "Negative", "Skipped" or "Positive",
the background color of the corresponding cell should be set
to red, yellow or green respectively. This currently works fine with the
Firefox browser but not with Internet Explorer (which indicate some error
when clicking on the image but no further information seems available).
If somebody knows how to fix that, we are interested (the
javascript code is quite simple (use view
source)).
Feedback to: Georges.Quenot@imag.fr