In the realm of computer vision, scale imbalance denotes a circumstance where the sizes of specific classes or objects within an image dataset vary significantly, with some being notably smaller than others. This situation presents a substantial challenge during the training of computer vision models. The model might develop a bias towards larger objects or classes, resulting in subpar performance when dealing with smaller ones.
Scale imbalance can emerge in diverse scenarios. For example, in satellite imagery, some features of interest could be relatively minute compared to others. Similarly, in medical imaging, certain anomalies may be of a smaller size. If not appropriately resolved, scale imbalance can give rise to low accuracy in image analysis and incorrect image classification.
Scale imbalance can be further categorized into two types: box-level scale imbalance and feature-level scale imbalance. Box-level scale imbalance pertains to the differences in the size of bounding boxes that enclose objects, while feature-level scale imbalance focuses on the variations in the scale of features extracted from the objects.