Reproduce “AD-Det: Boosting Object Detection in UAV Images with Focused Small Objects and Balanced Tail Classes” on OpenResearch

About this paper

Zhenteng Li, Sheng Lian, Dengfeng Pan, Youlin Wang, Wei Liu

The AD-Det framework enhances object detection in UAV imagery by simultaneously mitigating challenges from complex object scale variations and severe class imbalance through a coarse-to-fine strategy. It achieved new state-of-the-art average precision on the VisDrone and UAVDT datasets.

Key insights

• A synergistic, coarse-to-fine framework effectively addresses both small object detection and class imbalance simultaneously in UAV images, outperforming methods that tackle these issues in isolation.
• The Adaptive Small Object Enhancement (ASOE) module provides an unsupervised, plug-and-play mechanism to enhance small object detection by adaptively cropping subregions based on high-resolution feature map activations, without relying on additional density networks.
• The Dynamic Class-Balanced Copy-Paste (DCC) module introduces a novel object-level resampling strategy that ensures diversity through dynamic memory banks and optimizes pasting positions using context-aware search, balancing tail-class representation without introducing unrealistic artifacts.

Summary by alphaXiv.