% 28 * 28 for conv3
% 14 * 14 for conv4
% 7 * 7 for conv5
close all;
datas = load('datas5224_pure3.mat');
datas = datas.datas;
f = @(x,meanx,nSigma) exp(-1/2*(x-meanx)'*nSigma*(x-meanx));
for i = 1:length(datas)
landmark = datas(i).after_landmark;
imgpath2 = datas(i).imgpath2;
%imgpath = strrep(imgpath,'flickr','flickr224');
img = imread(imgpath2);
subplot(141);
imshow(img);
hold on;
for j = 1:size(landmark,1)
plot(landmark(j,1),landmark(j,2),'.r','markersize',15);
text(landmark(j,1),landmark(j,2),num2str(j),'color','w');
end
subplot(142);
heatmap = compute_heatmap([28,28],imgSize,[landmark(8,:);landmark(12,:);landmark(15,:)]);
imshow(heatmap,[]);
subplot(143);
heatmap = compute_heatmap([14,14],imgSize,[landmark(8,:);landmark(12,:);landmark(15,:)]);
imshow(heatmap,[]);
subplot(144);
heatmap = compute_heatmap([7,7],imgSize,[landmark(8,:);landmark(12,:);landmark(15,:)]);
imshow(heatmap,[]);
break;
end
function heatmap = compute_heatmap(imgsize,imgSize,landmarks)
% imgsize: the heatmap size for dst
% imgSize: the heatmap size for src, examples: 224 * 224 for CNN
% landmarks
landmarks = round(landmarks .* (imgsize ./ imgSize));
%tau = max((imgSize ./ imgsize)) * 0.1;
tau = min((imgsize ./ imgSize)) * 20; % change those values
Sigma = [tau^2 0 ; 0 tau^2]; % not use
nSigma = [1.0/(tau^2) 0;0 1.0/(tau^2)]; % inverse of Sigma
% note that: x and meanx's shape: [2,1], not [1,2]
f = @(x,meanx,nSigma) exp(-1/2*(x-meanx)'*nSigma*(x-meanx));
heatmap = zeros(imgsize(1),imgsize(2));
for i = 1:size(landmarks,1) % nlandmark
landmark = landmarks(i,:); % x,y
for j = 1:imgsize(2) % x: w
for k = 1:imgsize(1) % y: h
seat = [j,k]; % (x,y)
heatmap(k,j) = heatmap(k,j) + f(seat',landmark',nSigma);
end
end
end
activation_type = 'softmax';
if strcmp(activation_type,'sigmoid')
% pass Sigmoid activation funcion
heatmap = 1 ./ (1+exp(-heatmap));
elseif strcmp(activation_type,'softmax')
% pass softmax function
heatmap = exp(heatmap);
heatmap = heatmap ./ sum(heatmap(:));
end
end
produce gaussian heatmap for landmarks
最新推荐文章于 2021-12-22 21:47:03 发布