Training convolutional neural networks for person re-identification

Person ReID is the problem of matching people across many different camera views, also known as multi-camera tracking. This is an important area of research due to its usefulness in public security applications. Compared to other machine learning problems such as product searching, person ReID is an extremely challenging task, especially in real-world environments. This is due to variances in training data, including illumination condition, resolution, different viewpoint, and partial occlusion of individuals. Therefore, the objective of this project was to build a person ReID system that trains a neural network on many public datasets (e.g. Market1501, DukeMTMC-reID) to seek the most discriminative projections of the image features extracted from the previous neural network layers to adapt to the NTUCampus dataset. This project focuses on triplet-based deep similarity learning and domain adaptation. In this project, different combinations of algorithms were tested on many datasets to find a suitable combination of algorithms that will be good for domain adaptation. From the experiments, it was shown that there was not one combination of algorithms that performed the best for domain adaptation. Instead, the model to choose will depend on which dataset the model is going to be tested on. For testing on NTUCampus dataset, the models that were trained without softmax performed relatively better than the models that were trained with softmax. This can be a consideration when conducting future experiments.Bachelor of Engineering (Electrical and Electronic Engineering