Towards Governance-Oriented Low-Altitude Intelligence: A Management-Centric Multi-Modal Benchmark With Implicitly Coordinated Vision-Language Reasoning Framework

TL;DR

This work addresses governance-oriented reasoning in low-altitude urban scenes by shifting from exhaustive object detection to selective anomaly understanding. It introduces GovLA-10K, a multimodal dataset focusing on nine functionally salient governance targets, and GovLA-Reasoner, a lightweight feature adapter that implicitly coordinates visual grounding with LLM-based reasoning. The method achieves significant gains without fine-tuning detectors or LLMs and demonstrates robustness and deployment efficiency. Together, GovLA-10K and GovLA-Reasoner establish a practical foundation for governance-aware aerial intelligence and open avenues for future research in management-oriented vision–language systems.

Abstract

Low-altitude vision systems are becoming a critical infrastructure for smart city governance. However, existing object-centric perception paradigms and loosely coupled vision-language pipelines are still difficult to support management-oriented anomaly understanding required in real-world urban governance. To bridge this gap, we introduce GovLA-10K, the first management-oriented multi-modal benchmark for low-altitude intelligence, along with GovLA-Reasoner, a unified vision-language reasoning framework tailored for governance-aware aerial perception. Unlike existing studies that aim to exhaustively annotate all visible objects, GovLA-10K is deliberately designed around functionally salient targets that directly correspond to practical management needs, and further provides actionable management suggestions grounded in these observations. To effectively coordinate the fine-grained visual grounding with high-level contextual language reasoning, GovLA-Reasoner introduces an efficient feature adapter that implicitly coordinates discriminative representation sharing between the visual detector and the large language model (LLM). Extensive experiments show that our method significantly improves performance while avoiding the need of fine-tuning for any task-specific individual components. We believe our work offers a new perspective and foundation for future studies on management-aware low-altitude vision-language systems.

Figures (5)

• Figure 1: Annotation pipeline of GovLA-10K. Our GovLA-10K adopts a two-stage semi-automatic annotation pipeline. (a) Bounding box annotation stage: interest regions are manually annotated and cross-verified using a powerful detection model to ensure annotation accuracy. (b) Caption annotation stage: verified box annotations and original images are converted into structured, fine-grained prompts and fed into VLM to generate contextual captions. Peculiarly, all generated captions are also further reviewed by human annotators to maintain high annotation quality.
• Figure 2: Basic statistics of GovLA-10K. (a) Category-wise instance counts and proportions. (b) High-frequency word distribution in generated captions. (c) Numbers of images and bounding boxes in training and test sets. Please enlarge figure to capture more details.
• Figure 3: Distribution of image numbers across different caption lengths.
• Figure 4: Low-altitude framework overview of the existing pipeline (left) and our proposed GovLA-Reasoner (right). To address the information loss and error accumulation problems caused by the explicit prompt construction, we introduce an efficient adapter, which adaptively compresses and aggregates the discriminative features of grounding model to directly provide the robust feature representation to LLM. Note: snowflakes indicate frozen parameters, whereas flames indicate active (trainable) parameters.
• Figure 5: Detailed feature process pipeline of our designed Adapter.