Re-identify pedestrians with CNN.
A dataset created for person re-identification task, including 274 identities, 38934 bounding boxes, created by the lab of Professor Hua Yang of Shanghai Jiao Tong University.
A tool helps to clip bounding boxes from original images from video cameras. The code of the tool is here.
Train the model with a CNN modified from Caffenet, whose task is classification. The sizes of fc6 and fc7 are modified into 1024 in order to avoid overfitting. The size of fc8 is modified into 151, which is the number of identities in the training set.
Abandong fc8 and extract the vector of fc7 as the feature of an image. Save the features of query sequence and library sequence into a .mat file.
Adopt XQDA as metric learning method. Use CMC curve to show the performance. The code of evaluation is here.
- Python-2.7
- Caffe(for python)
- Numpy
- Pylab
Since the test set and training set are too large, so I uploaded them to MEGA. Get the deploy file, models and test set(4.51GB) by typing this command in the root directory:
sh download_model_proto_test.shGet the training set(11.65GB) by typing this command in the root directory (This step can be skipped if you just want to do the test):
sh download_train.shTrain:
python finetuning_RIDLT.pyYou have to change the directory of caffe or caffe-master in the code.
python evaluation_RIDLT.pyYou have to change the directory of caffe or caffe-master in the code.