Foreground attentive feature learning for person re-identification

A neural network architecture, comprising:

• a foreground attentive subnetwork, including:
• an encoder subnetwork configured to extract features from an RGB image, including passing the RBD image through two convolutional layers including learned filters, a rectified linear unit (ReLU), and a max pooling kernel;
• and
• a decoder subnetwork configured to formulate a binary mask of the RGB image foreground, including passing the extracted features through two deconvolutional layers including learned filters;
• a body part subnetwork configured to:
• concurrently with formulating the binary mask, averagely slice the extracted features into four equal sliced feature maps;
• for each of the four equal sliced feature maps:
• discriminatively learn feature representations of different body parts, including passing the sliced feature map two convolutional layers, a rectified linear unit (ReLU), and two max pool kernels;
• and
• learn a local feature map of a body part, including passing output of the two max pool kernels to a first fully connected layer, and passing output from the first fully connected layer to a second fully connected layer;
• a feature fusion subnetwork configured to:
• concatenate outputs from the first fully connected layers into a third fully connected layer;
• concatenate outputs from the second fully connected layers and the third fully connected layer into a normalization layer;
• and
• normalize feature vectors on a unit sphere space;
• and
• a symmetric triplet loss layer configured to learning final feature vectors.