|
|
| 面向低空无人机对地态势感知的样本级状态偏置校准方法 |
| Sample-Level State Bias Calibration for Scene Relationship Perception in Low-Altitude UAV Imagery |
| 投稿时间：2026-07-25 修订日期：2026-08-12 |
| DOI： |
| 中文关键词: 低空无人机 对地态势感知 场景图生成 长尾关系 样本级偏置校准 |
| English Keywords:low-altitude UAV ground situation perception scene graph generation long-tailed relation sample-level bias calibration |
| 基金项目:天基智能信息处理全国重点实验室基金项目 |
|
| 摘要点击次数: 90 |
| 全文下载次数: 0 |
| 中文摘要: |
| 针对低空无人机俯视场景中目标密集及关系类别长尾导致的高频关系预测偏置问题，提出了一种旨在提升结构化场景关系感知均衡性的校准方法。该方法采用冻结基础关系预测器的策略，通过融合候选目标对的联合区域视觉特征与关系分类前的模型内生状态，利用轻量偏置生成器预测样本级修正量，进而以残差方式校准原始关系分类得分。实验结果表明，该方法相较于Motif、PE-Net和RPCM等模型，在AUG数据集的关系分类任务上，综合评价总体预测能力和类别均衡能力的调和均衡指标H@100平均提高0.51%；在更具挑战的STAR数据集旋转框关系分类任务中，平均提高4.33%。该方法无需重新训练基础模型，可直接作为低空智能飞行视觉系统的插件式场景关系认知增强单元。 |
| English Summary: |
| To address the bias toward frequent relationships caused by dense object distributions and long-tailed relationship categories in low-altitude UAV top-view scenes, this paper proposes a calibration method for balanced scene relationship perception. The method freezes the backbone relationship predictor and employs a lightweight bias generator to estimate sample-specific residuals by jointly exploiting union-region visual features and pre-classification representations of candidate object pairs. The residuals are added to the original relationship scores for bias calibration. Compared with representative models including Motif, PE-Net, and RPCM, the proposed method improves the harmonic metric H@100 by an average of 0.51% on the AUG PredCls task and 4.33% on the more challenging STAR OBB PredCls task. Without retraining the backbone predictor, the proposed method can be directly deployed as a plug-in module for low-altitude aerial scene relationship perception. |
| 查看/发表评论 下载PDF阅读器 |
