Dynamic weighting label assignment for oriented object detection

Abstract

Oriented object detection has garnered significant attention for its broad applications in remote sensing image processing. Most oriented detectors perform dense predictions on a set of predefined anchors to generate oriented bounding boxes, where these anchors require classification (cls) and localization (loc) labels for detector training. Recent advancements in label assignment utilize the overall quality score of cls and loc predictions to determine positive and negative samples for each oriented object. However, these methods typically establish the overall quality score by assigning fixed weights to cls and loc quality scores. This approach may not be optimal, as fixed weights fail to dynamically balance cls and loc performance during model optimization, thereby constraining detection efficacy. Motivated by this observation, this paper proposes a Dynamic Weighting Label Assignment (DWLA) algorithm. DWLA dynamically adjusts the weights of individual quality scores based on the current model state to continuously balance cls and loc performance. Additionally, to mitigate the impact of unreliable predictions and achieve more stable training, this paper proposes a level-wise positive sample selection scheme and an object-adaptive scheme for constructing initial candidates of positive samples, respectively. Comprehensive experiments on the DOTA and UCAS-AOD datasets have validated the effectiveness of the proposed DWLA.