Extension of Intersection Over Union to Improve Small Object Detection in Few-Shot Regime
Pierre Le Jeune, Anissa Mokraoui · 2025
Intersection over Union (IoU) is a widely used criterion that quantifies the overlap between two bounding boxes. It plays a crucial role in object detection, serving both as a cost function for training detection models and as a criterion for evaluating their performance. The IoU value between the ground truth and predicted boxes determines whether a detection is considered accurate, based on a predefined threshold. However, this approach poses challenges when detecting small objects, especially in situations where annotated data is scarce, such as in few-shot learning scenarios. The scarcity of supervision hinders the learning of robust localization, which is especially detrimental for small objects. A small discrepancy of just a few pixels between the predicted and annotated bounding boxes can result in a false detection for small objects. To address these issues, we propose Scale-adaptive Intersection over Union (SIoU), a new controllable and adaptive similarity criterion that adjusts based on object size. First, SIoU helps to find a better balance between small and large objects during the training of few-shot detection methods, for which small objects are extremely problematic. Experiments on four distinct datasets show superior detection performance when using SIoU as a cost function. Second, by being more lenient with small objects, SIoU aligns more closely with human perception than IoU, making it a more suitable evaluation criterion.