Drone-PUR: From Visual Clue Discovery to Understanding and Reasoning in UAV Scenes

Title: Drone-PUR: From Visual Clue Discovery to Understanding and Reasoning in UAV Scenes

标题：Drone-PUR：从无人机场景中的视觉线索发现到理解与推理

Abstract: The reasoning capabilities of vision-language models (VLMs) can be limited by incomplete or inaccurate visual observations in the general domain. In unmanned aerial vehicle (UAV) scenes, small targets, subtle attributes, and changing viewpoints further complicate perception and understanding.

摘要：视觉语言模型（VLM）的推理能力可能会受到通用领域中不完整或不准确的视觉观察的限制。在无人机（UAV）场景中，微小的目标、细微的属性以及不断变化的视角进一步增加了感知和理解的复杂性。

Therefore, we introduce Drone-PUR, a dataset and benchmark organized around visual clue discovery, interpretation, and integration, corresponding to the three capability dimensions of perception, understanding, and reasoning. Reasoning samples are designed to require models to combine multiple visual clues to derive visually grounded conclusions.

因此，我们引入了 Drone-PUR，这是一个围绕视觉线索发现、解释和整合而组织的数据集和基准，对应于感知、理解和推理这三个能力维度。推理样本的设计要求模型结合多个视觉线索，以得出具有视觉依据的结论。

Its 11 tasks provide 175k training and 20k test samples spanning single images, image sequences, cross-modal images, and multiple viewpoints. All test samples undergo human review. Evaluating 22 general-purpose VLMs reveals persistent limitations in spatial judgment and cross-modal and cross-view reasoning.

其 11 项任务提供了 17.5 万个训练样本和 2 万个测试样本，涵盖了单幅图像、图像序列、跨模态图像和多视角图像。所有测试样本均经过人工审核。对 22 个通用 VLM 的评估揭示了它们在空间判断以及跨模态和跨视角推理方面存在的持续局限性。

Across models, stronger perception is associated with better reasoning, while single-task fine-tuning further reveals selective bidirectional transfer between these capabilities. To examine domain adaptation, we develop UAV-PUR from Qwen3-VL through staged fine-tuning with auxiliary supervision.

在不同模型中，更强的感知能力与更好的推理能力相关联，而单任务微调进一步揭示了这些能力之间存在选择性的双向迁移。为了检验领域适应性，我们通过带有辅助监督的分阶段微调，基于 Qwen3-VL 开发了 UAV-PUR。

Its 4B and 8B variants improve overall multiple-choice accuracy by 18.33 and 17.52 percentage points over their pretrained counterparts, respectively. The dataset and project resources are available at this https URL.

其 4B 和 8B 版本在多项选择题上的整体准确率分别比预训练模型提高了 18.33 和 17.52 个百分点。数据集和项目资源可在该 https 链接中获取。