gtolias / rmac Goto Github PK

% apply powerlaw
function x = powerlaw (x, a)
if a == 1, return; end
x = sign (x) .* abs(x)  .^ a;

Regarding to the powerlaw function above, "x" is the extracted feature vector and "a" is the original feature map of a different size, how is it possible to do abs(x) .^ a (element-wise power)?

Thanks!

I'd like to reproduce your research using your codes.
But, some errors prevent it.

I use 'matconvnet-1.0-beta25' and 'imagenet-caffe-alex.mat' that are downloaded from vlfeat.org, macOS 10.12.6, MATLAB2017a.

Error using bsxfun
Non-singleton dimensions of the two input arrays must match each other.

Error in apply_whiten (line 8)
x_ = bsxfun (@minus, x, xm);  % Subtract the mean

Error in test>@(y)vecpostproc(sum(vecpostproc(apply_whiten(y,mean_vectors_train,eigvec,eigval)),2))

Error in test (line 166)
            vecs_bestbox = cellfun(@(y) vecpostproc(sum(vecpostproc(apply_whiten(y, mean_vectors_train, eigvec, eigval)), 2)), vecs_bestbox, 'un', 0);

And, I got following error when I used 'matconvnet-1.0-beta15', macOS 10.12.6, MATLAB2017a.

Error using vl_simplenn (line 252)
Unknown layer type lrn

Error in rmac_regionvec (line 12)
	rnet = vl_simplenn(net, I);

Error in test>@(x)rmac_regionvec(imread(x),net,L)

Error in test (line 90)
            [vectors, conv3d] = cellfun(@(x) rmac_regionvec(imread(x), net, L), im_fn_test, 'un', 0);

What should I do for this issue?

first of all. thanks for your code and paper. I have a stupid question about the aggregation. why do you sum the maxpooling feature? It has any relationship with the integral Image ? Could you offer the paper name about this if possible.

i have read your paper, i have some question:

1. the rmac or mac is one of code for a image?
2. why you discard all fully connect layer?
3. you did not train vgg16 or alexnet?

ovr = 0.4; % desired overlap of neighboring regions
steps = [2 3 4 5 6 7]; % possible regions for the long dimension

W = size(X, 2);
H = size(X, 1);

w = min([W H]);
w2 = floor(w/2 -1);

b = (max(H, W)-w)./(steps-1);
[~, idx] = min(abs(((w.^2 - w.*b)./w.^2)-ovr)); % steps(idx) regions for long dimension

First of all, thank the author for sharing this code and article。Excuse me, why does the above code count like this? I donundefinedt understand.

Hello,
I read your paper of RMAC, thank you for sharing your code. It really offers great insights!

I am trying to reproducing some of the results from your paper, I am trying to draw the localization bounding box (region determined by AML) on the original image, while AML outputs the coordinates on the feature map (conv_5 layer). Can I ask how you relate the coordinates of the feature map to the coordinates of the original image, please?

Best
Thank you very much