Active Learning with Unfiltered Informativeness Technique for Object Detection

Contemporary Deep Learning models demand substantial volumes of data to effectively learn, creating a challenge given the difficulty of obtaining well-annotated data. Moreover, not all samples within a dataset are of equal significance to the learning process. Active Learning emerges as a solution to this issue by providing a structured framework for selecting the most instructive data within a dataset. This approach involves isolating a subset of the N most informative samples to train the algorithm effectively. In response to this challenge, we introduce Unfiltered Informativeness, a novel framework designed to assess the informativeness value of samples within a dataset. Our approach employs a trained detector to identify objects in a scene and subsequently computes the information gain of these objects. When multiple objects are present within a scene, we aggregate multiple scores into a single informative score. We systematically evaluate our approach against Random Sampling and other strategies.