Init

create files in the first time

data/computeOpticalFlow.m

@@ -0,0 +1,95 @@
+% Copyright (c) 2016 Niall McLaughlin, CSIT, Queen's University Belfast, UK
+% Contact: [email protected]
+% If you use this code please cite:
+% "Recurrent Convolutional Network for Video-based Person Re-Identification",
+% N McLaughlin, J Martinez Del Rincon, P Miller, 
+% IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016
+% 
+% This software is licensed for research and non-commercial use only.
+% 
+% The above copyright notice and this permission notice shall be included in
+% all copies or substantial portions of the Software.
+% 
+% THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+% IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+% FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+% AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+% LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+% OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+% THE SOFTWARE.
+
+%read an image sequence in the ilids video / PRID dataset
+%compute optical flow
+
+rootDir = fullfile('D:','person_re-id','data');
+
+for person = 1:319
+    disp(person)
+    for cam = 1:2
+
+        camName = {'cam_a','cam_b'};
+
+        dataDir = fullfile(rootDir,'i-LIDS-VID','sequences',['cam',num2str(cam)],['person',sprintf('%03i',person)]);
+        %dataDir = fullfile(rootDir,'PRID2011','multi_shot',camName{cam},['person_',sprintf('%04i',person)]);
+        files = dir(dataDir);
+
+        if exist(dataDir)
+
+            saveDir = fullfile(rootDir,'i-LIDS-VID-OF-HVP','sequences',['cam',num2str(cam)],['person',sprintf('%03i',person)]);
+            %saveDir = fullfile(rootDir,'PRID2011-OF-HVP','multi_shot',camName{cam},['person_',sprintf('%04i',person)]);
+            if ~exist(saveDir)
+                mkdir(saveDir);
+            end
+
+            seqFiles = {};
+            for f = 1:length(files)
+                if length(files(f).name) > 4 && ~isempty(findstr(files(f).name,'.png'))
+                    seqFiles = [seqFiles files(f).name];
+                end
+            end
+
+            optical = vision.OpticalFlow('Method','Lucas-Kanade','OutputValue', 'Horizontal and vertical components in complex form');
+
+            for f = 1:length(seqFiles)
+                seqImg = imread(fullfile(dataDir,seqFiles{f}));
+                optFlow = step(optical,double(rgb2gray(seqImg)));
+
+                %separate optFlow into mag and phase components
+                R = abs(optFlow);
+                theta = angle(optFlow);                
+
+                %threshold to remove pixels with large magnitude values
+                ofThreshold = 50;
+                R = min(R,ofThreshold);
+                R = max(R,-1*ofThreshold);                
+
+                %convert back to complex form
+                Z = R.*exp(1i*theta);                
+
+                H = imag(optFlow);
+                V = real(optFlow);
+                M = abs(optFlow);
+
+                H = H + 127;
+                V = V + 127;
+                M = M + 127;
+                P = theta + 127;
+
+                imgDims = size(seqImg);
+                tmpImg = zeros(imgDims);
+                tmpImg(:,:,1) = H;
+                tmpImg(:,:,2) = V;
+                tmpImg(:,:,3) = 0;
+
+                tmpImg(tmpImg < 0) = 0;
+                tmpImg(tmpImg > 255) = 255;
+
+                tmpImg = tmpImg ./ 255;            
+
+                %save optical flow image to file
+                saveFile = fullfile(saveDir,seqFiles{f});
+                imwrite(tmpImg,saveFile);
+            end
+        end
+    end
+end

datasets/datasetUtils.lua

@@ -0,0 +1,106 @@
+-- Copyright (c) 2016 Niall McLaughlin, CSIT, Queen's University Belfast, UK
+-- Contact: [email protected]
+-- If you use this code please cite:
+-- "Recurrent Convolutional Network for Video-based Person Re-Identification",
+-- N McLaughlin, J Martinez Del Rincon, P Miller, 
+-- IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016
+-- 
+-- This software is licensed for research and non-commercial use only.
+-- 
+-- The above copyright notice and this permission notice shall be included in
+-- all copies or substantial portions of the Software.
+-- 
+-- THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+-- IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+-- FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+-- AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+-- LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+-- OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+-- THE SOFTWARE.
+
+local dataset_utils = {}
+
+-- given the dataset, which consists of a table where t[x] contains the images for person x
+-- split the dataset into testing and training parts
+function dataset_utils.partitionDataset(nTotalPersons,testTrainSplit)
+    local splitPoint = torch.floor(nTotalPersons * testTrainSplit)
+    local inds = torch.randperm(nTotalPersons)
+
+    -- save the inds to a mat file
+    --mattorch.save('rnnInds.mat',inds)
+
+    trainInds = inds[{{1,splitPoint}}]
+    testInds = inds[{{splitPoint+1,nTotalPersons}}]
+
+    print('N train = ' .. trainInds:size(1))
+    print('N test  = ' .. testInds:size(1))
+
+    -- save the split to a file for later use
+    -- datasetSplit = {
+    --     trainInds = trainInds,
+    --     testInds = testInds,
+    -- }
+    -- torch.save('./trainedNets/dataSplit_PRID2011.th7',datasetSplit)
+
+    return trainInds,testInds
+end
+
+-- the dataset format is dataset[person][camera][nSeq][nCrop][FeatureVec]
+-- choose a pair of sequences from the same person
+function dataset_utils.getPosSample(dataset,trainInds,person,sampleSeqLen)
+
+    -- choose the camera, ilids video only has two, but change this for other datasets
+    local camA = 1
+    local camB = 2
+
+    local actualSampleSeqLen = sampleSeqLen
+    local nSeqA = dataset[trainInds[person]][camA]:size(1)
+    local nSeqB = dataset[trainInds[person]][camB]:size(1)
+
+    -- what to do if the sequence is shorter than the sampleSeqLen 
+    if nSeqA <= sampleSeqLen or nSeqB <= sampleSeqLen then
+        if nSeqA < nSeqB then
+            actualSampleSeqLen = nSeqA
+        else
+            actualSampleSeqLen = nSeqB
+        end
+    end
+
+    local startA = torch.floor(torch.rand(1)[1] * ((nSeqA - actualSampleSeqLen) + 1)) + 1    
+    local startB = torch.floor(torch.rand(1)[1] * ((nSeqB - actualSampleSeqLen) + 1)) + 1
+
+    return startA,startB,actualSampleSeqLen
+end
+
+-- the dataset format is dataset[person][camera][nSeq][nCrop][FeatureVec]
+-- choose a pair of sequences from different people
+function dataset_utils.getNegSample(dataset,trainInds,sampleSeqLen)
+
+    local permAllPersons = torch.randperm(trainInds:size(1))
+    local personA = permAllPersons[1]--torch.floor(torch.rand(1)[1] * 2) + 1
+    local personB = permAllPersons[2]--torch.floor(torch.rand(1)[1] * 2) + 1
+
+    -- choose the camera, ilids video only has two, but change this for other datasets
+    local camA = torch.floor(torch.rand(1)[1] * 2) + 1
+    local camB = torch.floor(torch.rand(1)[1] * 2) + 1
+
+    local actualSampleSeqLen = sampleSeqLen
+    local nSeqA = dataset[trainInds[personA]][camA]:size(1)
+    local nSeqB = dataset[trainInds[personB]][camB]:size(1)
+
+    -- what to do if the sequence is shorter than the sampleSeqLen 
+    if nSeqA <= sampleSeqLen or nSeqB <= sampleSeqLen then
+        if nSeqA < nSeqB then
+            actualSampleSeqLen = nSeqA
+        else
+            actualSampleSeqLen = nSeqB
+        end
+    end
+
+    local startA = torch.floor(torch.rand(1)[1] * ((nSeqA - actualSampleSeqLen) + 1)) + 1  
+    local startB = torch.floor(torch.rand(1)[1] * ((nSeqB - actualSampleSeqLen) + 1)) + 1
+
+    return personA,personB,camA,camB,startA,startB,actualSampleSeqLen
+end
+
+return dataset_utils