From: kmlee on
Hello to all

Recently, Hamed Masnadi-Shirazi released his Matlab source for Face
Detection using Adaboost as part of an assignment. I have taken the
code and modified it quite a bit. But importantly, optimized to run
30 times faster.

I have some interesting observations with which I'd like some
discussion / collaboration with.

I've tried looking at other boosting / machine learning boards, but
they look rather deserted.

For example, this code runs like the blazes in Matlab 6.1 but about
four times slower in 6.5.1.

When training to find the optimal threshold, at times it would select
the feature with the highest threshold. But there doesn't seem to be
anything wrong with the algorithm/data.

Please reply here or at the "cleaned up - ie remove all CAPS" version
of my email address.

Thanks in advance.
From: SeBy on
kmlee wrote:
>
>
> Hello to all
>
> Recently, Hamed Masnadi-Shirazi released his Matlab source for Face
> Detection using Adaboost as part of an assignment. I have taken the
> code and modified it quite a bit. But importantly, optimized to run
> 30 times faster.
>
> I have some interesting observations with which I'd like some
> discussion / collaboration with.
>
> I've tried looking at other boosting / machine learning boards, but
> they look rather deserted.
>
> For example, this code runs like the blazes in Matlab 6.1 but about
> four times slower in 6.5.1.
>
> When training to find the optimal threshold, at times it would
> select
> the feature with the highest threshold. But there doesn't seem to
> be
> anything wrong with the algorithm/data.
>
> Please reply here or at the "cleaned up - ie remove all CAPS"
> version
> of my email address.
>
> Thanks in advance.

Why not make a submission to FEX ? (File EXchange). Personaly I would
be interrested.

Regards,

Sýbastien
From: Vicente on
SeBy wrote:
> kmlee wrote:
>
>>
>>Hello to all
>>
>>Recently, Hamed Masnadi-Shirazi released his Matlab source for Face
>>Detection using Adaboost as part of an assignment. I have taken the
>>code and modified it quite a bit. But importantly, optimized to run
>>30 times faster.
>>
>>I have some interesting observations with which I'd like some
>>discussion / collaboration with.
>>
>>I've tried looking at other boosting / machine learning boards, but
>>they look rather deserted.
>>
>>For example, this code runs like the blazes in Matlab 6.1 but about
>>four times slower in 6.5.1.
>>
>>When training to find the optimal threshold, at times it would
>>select
>>the feature with the highest threshold. But there doesn't seem to
>>be
>>anything wrong with the algorithm/data.
>>
>>Please reply here or at the "cleaned up - ie remove all CAPS"
>>version
>>of my email address.
>>
>>Thanks in advance.
>
>
> Why not make a submission to FEX ? (File EXchange). Personaly I would
> be interrested.
>
> Regards,
>
> Sýbastien

me too.

Vicente Moreno.
From: kmlee on
Hi to all

SNIP
>
> Why not make a submission to FEX ? (File EXchange). Personaly I
> would
> be interrested.
>
> Regards,
>
> Sýbastien

I don't think the code is ready to be posted. I need more discussion
on the results and testing. A fresh perspective would help greatly!

Regards,
From: Steven Lord on

"kmlee" <kmleeAT(a)emailDOT.com> wrote in message
news:ef11a17.-1(a)webx.raydaftYaTP...
> Hello to all
>
> Recently, Hamed Masnadi-Shirazi released his Matlab source for Face
> Detection using Adaboost as part of an assignment. I have taken the
> code and modified it quite a bit. But importantly, optimized to run
> 30 times faster.
>
> I have some interesting observations with which I'd like some
> discussion / collaboration with.
>
> I've tried looking at other boosting / machine learning boards, but
> they look rather deserted.
>
> For example, this code runs like the blazes in Matlab 6.1 but about
> four times slower in 6.5.1.

Try profiling the code in both MATLAB 6.1 and 6.5.1, then look at what
section of the code took the longest in 6.5.1 or what slowed down the most
between the two MATLAB releases. If you need help figuring out a way to
improve the performance of that section, I'm sure the regular readers would
be happy to offer suggestions.

--
Steve Lord
slord(a)mathworks.com