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

Jiahe Wu¹Bing Cao¹Xingyue Guo^2,6Ruipu Zhao¹Beiming Yu¹Jiekang Feng^1,5Qinghua Hu¹Pengfei Zhu^1,3,4,5,†

¹Tianjin University²Nankai University³Southeast University⁴Low-Altitude Intelligence Laboratory, Xiong’an National Innovation Center⁵Xiong’an Guochuang Lantian Technology Company Ltd.⁶Nanyang Technological University

^†Corresponding author

Drone-PUR organizes visual clue discovery, semantic interpretation and evidence integration into perception, understanding and reasoning. A three-dimensional model comparison and a ten-task radar plot summarize evaluation.

11UAV tasks

3Capability dimensions

175,213Training samples

20,051Test samples

22General-purpose VLMs

Introduction

Small targets, subtle attributes and changing viewpoints make UAV imagery challenging for vision–language models. Reliable answers require models to discover task-relevant visual clues, interpret their semantic and spatial meaning, and connect multiple observations to support a conclusion. Drone-PUR is a dataset and benchmark organized around these three capability dimensions. Its 11 tasks span single images, image sequences, aligned RGB–thermal observations and synchronized UAV views. Training resources and complementary multiple-choice and open-ended evaluations support both capability assessment and domain adaptation.

Perception

Discover objects and extract categories, attributes, counts and locations.

Object Detection (OD)
Object Recognition (OR)
Object Counting (OC)
Multi-Image Grounding (MIG)

Understanding

Align visual clues with compound descriptions and spatial concepts.

Conditional Judgment (CJ)
Position Spatial Judgment (PSJ)
Visual Grounding (VG)

Reasoning

Integrate multiple visual clues to infer explanations and situations.

Causal Reasoning (CR)
Situational Inference (SI)
Multimodal Reasoning (MMR)
Multi-view Reasoning (MVR)

Samples

Multi-view Reasoning

Joint reasoning

QUESTION

CORRECT ANSWER

Drone-PUR Benchmark

Drone-PUR brings together 168K images and 620K bounding-box annotations. Around 90.2% of task-relevant boxes occupy less than 1% of an image, reflecting the fine-grained demands of UAV observation. Training and test splits are disjoint at the image and video levels. All test samples are human-reviewed. The Position Spatial Judgment task is entirely human-annotated and covers vehicle heading, pedestrian heading, turning direction, object-relative position and vertical-level comparison. Evaluation combines ten multiple-choice tasks with direct detection, grounding and counting.

Eleven Drone-PUR tasks grouped into perception, understanding and reasoning, with original example images, questions and reference answers. — The 11-task overview

Image sources, task distributions, proportions of human-authored or revised samples, and normalized region areas. — **Statistical overview of the Drone-PUR dataset.** (a) Image sources by task. (b–c) Task distributions in the benchmark and training sets, respectively. (d) Proportions of human-authored or revised samples by task and split. (e) The distribution of region areas normalized by image area.

Drone-PUR integrates images and annotations from eight UAV datasets previously released by our team: VisDrone, DroneVehicle, MDMT, MDOT, DroneRGBT, DroneCrowd, AnimalDrone, and DroneBird. We additionally introduce 66,940 newly collected images, extending coverage to urban, rural, industrial, forest, maritime, and anomalous scenes. The newly collected data were captured at heights of 8–400 m. The RGB–TIR pairs in the newly collected data are temporally synchronized but exhibit spatial offsets.

UAV-PUR: a unified adaptation baseline.

Built on Qwen3-VL at the 4B and 8B scales, UAV-PUR combines stage-wise fine-tuning with geometric and numerical auxiliary supervision. All tasks retain the original language-generation interface at inference.

01
Perception adaptation
Strengthen visual clue discovery through perception tasks.
02
Joint capability learning
Train across all three dimensions with eligible auxiliary supervision.
03
Final consolidation
Continue with language-generation supervision alone.

UAV-PUR uses training-only grounding, counting and detection heads during the first two stages, and the original language modeling head at inference.

Model Performance

TOP 10

Top 10 models ranked by the sample-weighted overall accuracy reported in Table 1.
Model	OR	OC	MIG	CJ	PSJ	VG	CR	SI	MMR	MVR	Avg.
UAV-PUR-8B	95.69	60.19	73.33	83.68	50.86	68.92	72.36	70.64	66.72	70.33	70.26
UAV-PUR-4B	95.46	59.23	72.63	82.68	50.28	69.51	71.68	70.18	65.73	69.01	69.56
Gemini-3.5-flash	93.67	39.01	50.91	83.92	36.34	73.70	75.56	78.66	68.96	69.61	66.09
Qwen3.5-plus	91.76	39.07	49.49	80.25	34.94	74.21	74.98	74.28	50.06	42.46	58.80
Qwen3-VL-235B	89.36	32.99	47.88	83.48	36.34	73.99	68.77	71.20	50.17	39.10	57.72
GPT-5.4	88.91	32.20	51.72	83.84	31.30	64.07	68.09	71.76	49.80	34.29	55.87
Qwen3-VL-32B	88.07	29.70	44.04	81.89	33.50	69.58	68.57	70.18	48.64	35.74	55.45
Qwen3.5-9B	90.03	31.69	40.91	81.25	32.59	69.07	65.76	66.82	48.08	39.40	55.27
DeepSeek-V4.1-flash	86.27	34.58	37.37	79.78	33.10	63.78	62.75	71.67	43.24	34.47	53.22
Qwen3-VL-8B	84.09	30.83	40.20	78.70	36.41	65.32	64.50	68.31	42.05	32.37	52.74

BibTeX

@misc{wu2026dronepur,
  title  = {Drone-PUR: From Visual Clue Discovery to
            Understanding and Reasoning in UAV Scenes},
  author = {Jiahe Wu and Bing Cao and Xingyue Guo and Ruipu Zhao
            and Beiming Yu and Jiekang Feng and Qinghua Hu and Pengfei Zhu},
  year   = {2026},
  url    = {https://drone-pur.github.io/}
}

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