Gaussian similarity-based adaptive dynamic label assignment for tiny object detection

Abstract

Benefiting from the advanced deep learning techniques, significant achievements have been made in generic object detection. Tiny object detection (TOD) is a challenging task in computer vision due to the low resolution, insufficient geometric cues, and high noise levels. A recent trend for detectors is introducing more granular label assignment strategies to provide promising supervision information for classification and regression. However, most previous Intersection-Over-Union (IoU) based methods suffer from two main drawbacks, including (1) low tolerance of IoU for bounding box deviations in tiny objects and (2) deficient guidance for optimization caused by inter-sample and intra-sample imbalance. We propose two novel components to address these problems: the Gaussian probabilistic distribution-based fuzzy similarity metric (GPM) and the adaptive dynamic anchor mining strategy (ADAS). GPM aims to address the issue of inaccurate similarity measurement between small bounding boxes and pre-defined anchors, providing a more accurate basis for label assignment. ADAS adopts a dynamically adjusted strategy for label assignment to address the distribution bias between positive and negative samples, ensuring that the label assignment is consistent with the distribution of objects in the image. Extensive experiments are conducted on AI-TODv2 and other tiny object detection datasets to evaluate the proposed ADAS-GPM method’s performance. The results demonstrate that incorporating ADAS-GPM into an anchor-based object detector yields significant outperformance over state-of-the-art methods on the challenging AI-TODv2 benchmark. The proposed ADAS-GPM method exhibits good results, clearly demonstrating its validity and potential.