Link to refer to this document:
Abstract
This paper presents a method to select a diverse subset of images from vehicle camera using unsupervised learning. The proposed approach addresses the costly process of manual data labeling by creating a subset of the most valuable datapoints, those with the greatest diversity. The method includes transforming video frames into n-dimensional vectors using contrastive unsupervised learning, regularizing the representation vectors using UMAP, and downsampling by selecting only the furthest vectors. The proposed method is subjectively evaluated on real-life dashcam videos. A web application is built using a fully custom made architecture, where the REST requests are handled by Flask and the worker nodes are handled with Celery. The priority is to handle multiple requests and make use of multiple worker nodes. The web application provides an easy-to-use interface for researchers and practitioners to select diverse images from vehicle cameras for training neural networks.
