To find the wall, I would invert the image, do a dilation and then & the original and the processed images. It doesn't really make a difference, but it makes more intuitive sense to me. The optimal size and shape of your SE depends on your application, but usually it doesn't matter very much.
However, I would suggest you limit the number of operations, so don't do 4 erosions if you actually only want 1.
im=imread('cameraman.tif');
L=bwlabel(im<=20);
im=L==1;%pick biggest blob
r=4;
[X,Y]=ndgrid(-r:r);
SE=hypot(X,Y)<=r;
im_shell_erode=~imerode(im,SE) & im;
im_shell_dilate=imdilate(~im,SE) & im;
im_rgb=double(im);
im_rgb(:,:,2)=im_shell_erode;
im_rgb(:,:,3)=im_shell_dilate;
figure(1),clf(1)%only use clf in debugging
imshow(im_rgb)
title('R=original, G=eroded, B=dilated')
