Boosting Tiny Object Detection in Complex Backgrounds Through Deep Multi-Instance Learning

Our research paper presents a groundbreaking computer vision network architecture that effectively detects small objects by integrating Object detection and multi-instance learning (MIL). Conventional MIL models often underperform when the assumption of negative bags containing solely negative instances and positive bags containing at least one positive instance is violated, which commonly occurs during MIL for object detection. Additionally, detecting small objects using MIL poses challenges such as false positives and imprecise categorizations due to factors like varying scale and complex backgrounds. To address these challenges, our model incorporates two Recursive Feature Pyramid Networks (RFPNs) and employs the generation of exclusive negative bags. Through extensive experimentation, we demonstrate that our model surpasses the performance of state-of-the-art models for detecting tiny objects. Moreover, our work represents the first empirical study on detecting small objects from complex backgrounds using (multi-instance learning) MIL. Additionally, our solution improves the accuracy and speed of multi-stage methods by incorporating multiple architectural enhancements.