Progressive unsupervised learning: the next step in re-identification deep learning

What is it about?

Person re-identification (re-ID) is a method that helps identify the same person across different cameras. It is a notable application of machine learning. Re-ID can tell us whether a person appears in another camera after being captured by one earlier. Re-ID methods based on deep learning (DL) have dominated literature and become the gold standard. Despite this, the accuracy of DL re-ID drops outside the bounds of the training set. This means it is hard for it to be applied to new and untested data. The common solution to such issues in machine learning is to resort to unsupervised learning. In this work, researchers have designed a method for unsupervised deep representation learning for re-ID. Named “progressive unsupervised learning” (PUL), the method is an iterative process that can find better representations without labeled training data. It can adapt to an unlabeled dataset and is easy to implement.

Featured Image

Photo by Nick Loggie on Unsplash

Photo by Nick Loggie on Unsplash

Why is it important?

Using untested datasets, re-ID could identify and classify previously unseen types of data. This could lead to easy adjustment to new locations. It could also lead to novel applications such as using re-ID as a search problem aiming to retrieve the most relevant images from any data. The more the method is used, the better it becomes. As its uses grow, the importance of re-ID will grow with it. During that process, new applications will emerge for the method. KEY TAKEAWAY: A simple method for person re-ID is designed based on an iterative approach and fine tuning the CNN. Sample selection is a key component to its success. The proposed method produces CNN models with high discriminative ability. It presents a huge leap forward in re-ID usage.