People re-identification using deep appearance, feature and attribute learning

Person Re-Identification (Re-ID) is the act of matching one or more query images of an individual with images of the same individual in a gallery set. We propose various methods to improve Re-ID performance via foreground modelling, skeleton prediction and attribute detection. Foreground modelling is an important preprocessing step in Re-ID, allowing more representative features to be extracted. We propose two foreground modelling methods which learn a mapping between a set of training images and skeleton keypoints. The first utilises Partial Least Squares (PLS) regression to learn a mapping between Histogram of Oriented Gradients (HOG) features extracted from person images, and skeleton keypoints. The second instead learns the mapping using a deep convolutional neural network (CNN). Using a CNN has been shown to generalise better, particularly for unusual pedestrian poses. We then utilise the predicted skeleton to generate a binary mask, separating the foreground from the background. This is useful for weighting image features extracted from foreground areas higher than those extracted from background areas. We apply this weighting during the feature extraction stage to increase matching rates. The predicted skeleton can be used to divide a pedestrian image into multiple parts, such as head and torso. We propose using the divided images as input to an attribute prediction network. We then use this network to generate robust feature descriptors, and demonstrate competitive Re-ID matching rates. We evaluate on a number of dfferent Re-ID data sets, each possessing significant variations in visual characteristics. We validate our proposals by measuring the rank-n score, which is equivalent to the percentage of identities correctly predicted within n attempts. We evaluate our skeleton prediction network using root mean square error (RMSE), and our attribute prediction network using accuracy. Experiments demonstrate that our proposed methods can supplement traditional Re-ID approaches to increase rank-n matching rates.