Adaptive Training Strategies for Small Object Detection Using Anchor-Based Detectors

Small object detection is a crucial task in computer vision due to its wide range of real-world applications. Detecting small objects accurately and efficiently remains a challenging task due to the reduced size of the objects, low contrast to their surroundings, and potential occlusions. To tackle this issue, we proposed a method for detecting small objects in object detection tasks, including a new strategy for balancing positive and negative samples, a loss function that adapts the weight of detection losses according to object size, and an anchor mechanism that accommodates objects with diverse sizes and aspect ratios. The experimental data substantiates that our method has achieved a 12.9% increase in average accuracy for small objects on the COCO dataset, compared to the baseline.