Person Reidentification and Classification

This work will follow the state-of-the-art method, named Joint Recurrent Learning, or JRL. JRL model consists of a convolutional neural network for feature extraction followed by a recurrent part organized as encoder-decoder. Image features are sliced horizontally into regions to create a sequence, while attributes are also ordered in a particular way. So the problem is reformulated as sequence-to-sequence translation. Besides, decoder is equipped with an attention mechanism to focus on a proper image region while predicting a particular attribute. The contributions of this work include implementation of the original JRL as well as exploring possibilities of its alteration with following modifications: *replacing basic convolutional network for feature extraction; *choosing other type of recurrent unit in the encoder-decoder part; *changing the strategy of extracting image features; *changing the method of aggregating predictions in the ensemble of models (a few - model are trained) ; *adding augmentation while training basic convolutional network. For each modified JRL quality is evaluated on dataset PETA to provide comparative results. As the result, it is shown that each modification results in a prediction quality gain with respect to the original model. The most effective strategy is to change the method of aggregating predictions in the ensemble. Modified JRL model enables to improve pedestrian attribute classification in case of low resolution, poor quality, inaccurate color rendition of surveillance images along with variety of attribute appearance and localization. Keywords: recurrent neural network, convolutional neural network, multi-label classification, pedestrian attribute classification.