Revisiting Visual Pattern Mining

Abstract

With the progress in deep learning based methods, visual pattern mining has seen a significant improvement in extracting visual patterns in the form of mid-level elements and using these patterns for object recognition tasks. The problem with the previous approaches is that they are fully supervised and requires a large amount of labelled data for pattern mining. But how to make it work when there is little or no labelled data? In this work, we propose an unsupervised pattern mining algorithm which works very well given a large unlabelled dataset. We further extend it to show how it also adapts to include labelled data as well and thus, is able to extract information from both labelled and unlabelled data together. This property makes it very useful for low-shot recognition tasks where the labelled data is present in very small quantities and there is an abundance of unlabelled data. In this work we show the effectiveness of our pattern mining algorithm on the task of low-shot fine grained recognition and image labelling. We show that our unsupervised mining algorithm is able to detect fine grained patterns of good quality even without using any labels and if given a few labelled images there is a significant improvement in quality and diversity of patterns. We also show the ability of our approach in labelling more images from the large unlabelled pool and adding them iteratively to the labelled set in a semi-supervised learning based approach. Our method performs much better than the baselines which include previous state of the art approaches to fine grained recognition.