Open Warrior

I’ve spent most of the afternoon trying to figure out how I could use MEncoder to record video and audio from my webcam and kept running into a “floating point exception” problem. In scouring the internet I realised that many people seem to be having the exact same problem. It turns out that the problem is due to a bug in MEncoder and is simply fixed by installing the latest version.

I use Kubuntu Jaunty so the problem for me lay in the version of Mencoder that comes packaged with it. To fix the problem all I did was follow these instruction for adding a thrid party repository for newer builds of MPlayer/MEncoder and then installed the latest MEncoder using apt-get.

sudo apt-get install mplayer mencoder

Thereafter I was able to capture video and audio from my webcam using the following command:

mencoder tv:// -tv driver=v4l2:width=320:height=240:device=/dev/video0:forceaudio:adevice=/dev/dsp -ovc lavc -oac mp3lame -lameopts cbr:br=64:mode=3 -o webcam.avi

A few weeks ago I wrote a post which introduced the BOLD Project. Well, a lot has happened since then and this post gives an overview of the translation system which I am building.

BOLD Translator Overview

The translator is split into three parts:

1. The preprocessor
2. The user input
3. The matcher

The Preprocessor

The preprocessor is called as soon as an image is inserted into the repository. The preprocessor works by first segmenting the Bushman words in the dictionary. It does this by exploiting the known fact that every Bushman word on a page is underlined by a solid black line. Once the Bushman words have been segmented, specific features are extracted from them and these features are stored in inverted files. Once the features have been extracted then the orifinal image, the segmented words and the inverted files are all stored in the repository.

The User Input

The user who is accessing the Bleek & Lloyd notebooks uses a tool to select a specific word on a page which then becomes known as the key. The same features which were extracted from each word in the preprocessor are extracted from the key. These features along with the key image will be used later for matching.

The Matcher

The matcher starts by taking the features belonging to the key and finding images with the same features in the inverted files. For each feature match, the score of the image which matched increases. At the end of all the feature comparisons, the images with the highest scores are returned. At this stage there may be some images with the same or similar scores, so to resolve this clash the matcher performs a more intensive comparison between the key and the images with the highest score. Based on the result of this comparison, the most likely match is returned.

So that’s how the BOLD Translator works. Ultimately it is a framework which means that it will be designed such that anyone can adapt it and make use of it by plugging their own algorithms into each of the specific parts. In the next day or two I will blog about the actual work that has been done on the system up to now as well as show some of the results that the translator returns at this point.