CityToolVQA improves urban 3D question answering with geometric tools
CityToolVQA: Tool-Augmented Visual Question Answering for 3D Spatial Cognition in Urban Low-Altitude Environments
Computer Vision and Pattern Recognition
Summary
Answering detailed questions about city scenes from images is hard for AI models that only use regular photos. The authors split these questions into ones about what things are and ones about measurements like size or distance. For measurement questions, they add a special tool that calculates 3D details using depth and geometry. This combination helps AI answer such questions much better, showing that using extra 3D tools helps where just looking at pictures falls short.
What this means in practice
- •For urban planners: Use combined AI and geometric tools to better evaluate spatial queries about city infrastructure from low-altitude imagery.
- •For drone operation teams: Improve accurate answers about object sizes and distances in urban drone footage by integrating 3D geometry tools with vision models.
Authors
Boao Yu, Yingzhen Nie, Yue Hu, Zhengqiu Zhu, Rusheng Ju
Abstract
CityToolVQA addresses the weak performance of Vision-Language Models (VLMs) on quantitative tasks in urban low-altitude visual question answering. We divide the seven tasks into qualitative and quantitative groups: qualitative questions are answered directly by the VLM, whereas quantitative questions are handled by an external visual-geometric toolchain that performs object grounding, segmentation, depth back-projection, and spatial computation. The toolchain can be attached to different VLMs in a zero-shot manner; a Depth-Assisted Prompt Inference (DAPI) fallback is triggered when the main-chain detection is invalid or unreliable, and CityToolVQA-SFT adapts the 8B backbone to tool-conditioned inputs. On the 73,324-question Open3D-VQA-v2 test set, CityToolVQA-SFT (Qwen3-VL-8B) reaches 67.6% overall accuracy, and attaching the toolchain to ten open-source VLMs improves quantitative-task accuracy by 12.7-36.9 percentage points. These results indicate that externalizing explicit 3D geometric computation effectively complements the limited ability of RGB-only VLMs to estimate metric distances and object sizes.