Help on Gesture recognition code?

I'm doing a project on gesture recognition using Neural Networks. The training Network is RBF net, and the net is saved as gesture1net in an m-file train.m . However, when I load the network, I get an error message that says "undefined variable or function gesture1net". Can anyone tell me why that is? Here's part of the code:

1) train.m

% This program is used to make the neural network for the identification of the gesture1.

clc;

clear all;

% Load the sample images (of different poses) to make net for gesture1

img1=imread('firsgesture1.jpg');

img2=imread('firsgesture2.jpg');

img3=imread('firsgesture3.jpg');

% Crop the sample images

imgcpd1=imcrop(img1);

imgcpd2=imcrop(img2);

imgcpd3=imcrop(img3);

% Train the network. p is the input vector and t is the target vector

p=[imgcpd1,imgcpd2,imgcpd3];

t=[imgcpd1,imgcpd1,imgcpd1];

p=double(p);

t=double(t);

net=newrb(p,t,.001,.1,1000,25); %train using RBF net

save('gesture1net','net'); %save the network

2) trackrecog.m

% MATLAB program for real time hand tracking and gesture identification

clc;

clear all;

close all;

train;

closepreview;

vid = videoinput('winvideo', 1, 'YUY2_640x480'); %specify the video adaptor

src = getselectedsource(vid);

vid.ReturnedColorspace = 'grayscale'; %define the color format to GRAYSCALE

vid.FramesPerTrigger = 1;

preview(vid); %preview the video object

load gesture1net; %load the network for each gesture

imgref1=imread('refimg1.jpg'); %loading the reference image for each gesture

imgcrpd1=imcrop(imgref1); %crop the reference images

imgcrpd1=double(imgcrpd1);

imgtrained1=sim(gesture1net,imgcrpd1) ; %train the cropped reference images_

Thank You!

Answer 1

[Walter Roberson]

Do you get the error message on the load statement? Or on something you do right after the load? With that form of "load" the variable that would be created would be named "net".

Also, your invocation of train should be only

train

and not

train.m

[Pruthvi]

Oh, I'm so sorry, the train.m was a typo! The error is in the last line,

imgtrained1=sim(gesture1net,imgcrpd1);

You're right, it's after the load statement. Sorry for not being clear earlier!

[Walter Roberson]

The variable you are saving is named "net", not "guesture1net". load() that does not have any output arguments restores variables under their original name.

You can change your train.m to have

gesture1name = newrb(p,t,.001,.1,1000,25); %train using RBF net

save('gesture1net','gesture1net'); %save the network

[Image Analyst]

You're calling a function that is supposed to return Pi but in your code, because of some conditions on if statements, for loops, or while loops, or case statements, the code that is supposed to set Pi never gets called.

[Walter Roberson]

Did the error come out with the original network or after it was loaded?

Your "train" routine is a script rather than a function, so any variables it creates are still in the workspace. You would not need to save() or load() to access the network.

[Pruthvi]

I converted the imgcrpd1 to double and got rid of the previous error. And then I scrapped the save() and load() functions and got this error :

??? Error using ==> vertcat
CAT arguments dimensions are not consistent.

Error in ==> calcpd at 81
          pd = [pd; Pc{j,ts-d}];

Error in ==> network.sim at 186
Pd = calcpd(net,TS,Q,Pc);

Error in ==> trackrecog at 37
imgtrained1=sim(net1,imgcrpd1); %train the cropped reference images

[Walter Roberson]

Are all of the gesture images the same size? Are you cropping them to exactly the same size?

[Image Analyst]

You need to learn how to use the debugger. Once you do that you can figure out what the dimensions of pd are and what the dimensions of whatever array is inside the cell Pc{j,ts-d}, and you can see that the numbers of columns don't match. I don't know why - you have to figure out why you are wanting to vertically stack arrays that don't have the same number of columns. You can't tack on a 10 by 10 array onto the bottom of a 2 by 2 array - their numbers of columns don't match.

[Pruthvi]

Thanks a million for your help! I found out where I went wrong, and the program works now! Thanks again!

[Greg Heath]

1. train is a MATLAB function in the NNTBX

 help train
 doc train

I recommend using another name.

2. I am very suspicious of the inputs in

 net = newrb(p,t,.001,.1,1000,25); %train using RBF net

H = 1000 indicates to me that the net is probably overfit, overtrained, and, probably, unsuccessful on nontraining data.

3. Will you please post the results of the command

 whos

so that I can see the dimensions of your variables?

Hope this helps.

Greg

[Pruthvi]

Well, the performance doesn't meet the goal even after H=1000. The MSE is 126.07. But you're right, the network isn't able to classify other images, and the program doesn't work very well. Here's our code. Could you suggest any changes to make it better?