Datasets:

Henard224
/

Drone-PUR-Benchmark

question_id string	images list	question string	correct_answer string
CJ_merged_000001	[ { "path": "images/Dronemillion/img_008024.jpg" } ]	Are there trains present on both levels of the elevated railway, with at least one train on each track?	No
CJ_merged_000002	[ { "path": "images/Dronemillion/img_008030.jpg" } ]	Is the small red truck beneath the elevated railway parked tightly against the building wall with its doors closed, and is there no visible barrier or marked pedestrian sidewalk between the vehicle and the wall?	No
CJ_merged_000004	[ { "path": "images/Dronemillion/img_008039.jpg" } ]	To the right of the flatbed truck with a green-and-black cab parked on the left side of the elevated tracks, are there two sedans with their front hoods open?	Yes
CJ_merged_000005	[ { "path": "images/Dronemillion/img_008045.jpg" } ]	Is the greenery strip between the elevated railway and the highway densely covered with healthy green foliage, without any visibly bare ground patches?	No
CJ_merged_000006	[ { "path": "images/Dronemillion/img_008049.jpg" } ]	Among the cars parked along the road immediately to the left of the beige-roofed building, are there blue, red, and orange cars?	Yes
CJ_merged_000007	[ { "path": "images/Dronemillion/img_008056.jpg" } ]	On the main multi-lane road on the left, are vehicles traveling in both directions with each lane kept mostly clear of stopped or illegally parked cars?	No
CJ_merged_000009	[ { "path": "images/Dronemillion/img_008068.jpg" } ]	In the current scene, can a gas station, residential apartment buildings, and a material storage and transport factory all be seen at the same time?	Yes
CJ_merged_000010	[ { "path": "images/Dronemillion/img_008074.jpg" } ]	In the parking area between the factory and the elevated tracks, are all cars parked within clearly marked spaces and arranged in a neat single row, with no vehicles skewed or occupying spaces side by side?	No
CJ_merged_000011	[ { "path": "images/Dronemillion/img_008080.jpg" } ]	Within this greenery strip, are there visible walking paths with people using them and benches placed alongside the paths?	No
CJ_merged_000014	[ { "path": "images/Dronemillion/img_008093.jpg" } ]	In the area beside the yellow container building, is there at least one car parked close to the building while temporary barriers are positioned nearby to control vehicle access?	Yes
CJ_merged_000015	[ { "path": "images/Dronemillion/img_008099.jpg" } ]	Within the parking area beside the tracks, are all parked cars positioned fully within marked spaces and facing the railway lines in an orderly, uniform orientation?	No
CJ_merged_000016	[ { "path": "images/Dronemillion/img_008103.jpg" } ]	Are any railway maintenance workers operating on the tracks, and can a track junction or switching point be seen in the image?	No
CJ_merged_000017	[ { "path": "images/Dronemillion/img_008109.jpg" } ]	Is the railway station area free of pedestrians while at least one train is present at the covered platform and multiple cars are parked along the adjacent curved access road?	Yes
CJ_merged_000018	[ { "path": "images/Dronemillion/img_008113.jpg" } ]	Are there multiple sets of tracks on the elevated railway, and does the image show both a track junction or switching point and an area closed for maintenance?	No
CJ_merged_000045	[ { "path": "images/Dronemillion/img_008123.jpg" } ]	Despite the heavy fog, are several vehicles on the central road still maintaining enough distance from one another to avoid clustering together?	Yes
CJ_merged_000054	[ { "path": "images/Dronemillion/img_008130.jpg" } ]	Under dense fog conditions, are multiple vehicles on the elevated road still keeping clearly visible gaps between one another?	Yes
CJ_merged_000055	[ { "path": "images/Dronemillion/img_008132.jpg" } ]	Can any marked sports facility be clearly observed in the fog-covered scene, including visible boundary lines and goals or nets?	No
CJ_merged_000056	[ { "path": "images/Dronemillion/img_008139.jpg" } ]	Are there any vehicles parked vertically (with their fronts facing either downward or upward) in the parking area at the lower-right corner?	Yes
CJ_merged_000057	[ { "path": "images/Dronemillion/img_008146.jpg" } ]	In the curved parking area beside the blue-roofed building, are all vehicles neatly parked in marked spaces with their fronts facing the building?	No
CJ_merged_000058	[ { "path": "images/Dronemillion/img_008152.jpg" } ]	In the parking area along the bottom edge, are most cars parked within marked parking spaces, roughly parallel to the curb line, and not blocking the internal lane?	Yes
CJ_merged_000059	[ { "path": "images/Dronemillion/img_008162.jpg" } ]	In the row of parking spaces between the landscaped strip on the left side of the central vertical road and the high-rise building on the left side of the image, are there any red and blue cars parked there?	No
CJ_merged_000060	[ { "path": "images/Dronemillion/img_008171.jpg" } ]	At the central multi-lane intersection, are there no vehicles on the crosswalk, and is there a bus with a blue roof to the right of the intersection?	No
CJ_merged_000061	[ { "path": "images/Dronemillion/img_008182.jpg" } ]	Is the image foggy with low visibility, and are many pedestrians walking on the crosswalk?	No
CJ_merged_000062	[ { "path": "images/Dronemillion/img_008188.jpg" } ]	In the parking lot directly below the building on the left side of the image, are most parked cars neatly positioned in parking spaces, generally facing the building entrance, while their bodies are randomly tilted?	No
CJ_merged_000063	[ { "path": "images/Dronemillion/img_008199.jpg" } ]	Is there no snow visible anywhere on the rooftops or ground in the image anymore?	No
CJ_merged_000064	[ { "path": "images/Dronemillion/img_008210.jpg" } ]	In the parking area along the right side of the main road, are cars parked in marked bays while still leaving a clear lane for vehicles to drive through?	Yes
CJ_merged_000065	[ { "path": "images/Dronemillion/img_008220.jpg" } ]	Are the vehicles parked on both sides of the main road arranged within clearly marked parking spaces, leaving a continuous and unobstructed emergency lane?	No
CJ_merged_000073	[ { "path": "images/Dronemillion/img_008222.jpg" } ]	Are all vehicles waiting behind the stop lines at the central intersection while the pedestrian crosswalks remain completely unobstructed?	No
CJ_merged_000075	[ { "path": "images/Dronemillion/img_008235.jpg" } ]	Within the central intersection, are all vehicles aligned strictly within lane markings while simultaneously yielding clear space around the pedestrian crosswalks?	No
CJ_merged_000076	[ { "path": "images/Dronemillion/img_008244.jpg" } ]	Under low-visibility conditions, are vehicles from all four directions leading to the central intersection generally aligned within their lanes and stopped before the crosswalks rather than blocking them?	Yes
CJ_merged_000077	[ { "path": "images/Dronemillion/img_008254.jpg" } ]	Among the row of parked vehicles below the dental hospital, are there any white, gray, and pink cars?	Yes
CJ_merged_000078	[ { "path": "images/Dronemillion/img_008260.jpg" } ]	Does the scene show both heavy smog reducing visibility and vehicles still actively using the intersection in multiple directions?	Yes
CJ_merged_000079	[ { "path": "images/Dronemillion/img_008265.jpg" } ]	At the central intersection, are multiple vehicles present within the junction area while the pedestrian crosswalk markings in all four directions remain clearly unobstructed by vehicles?	Yes
CJ_merged_000080	[ { "path": "images/Dronemillion/img_008272.jpg" } ]	At the central intersection, are moving vehicles generally aligned within the lane markings while there is still a vehicle on the crosswalk?	No
CJ_merged_000085	[ { "path": "images/Dronemillion/img_008287.jpg" } ]	Are the vehicles parked in the open area at the lower right arranged in different orientations while all aligned with clearly visible parking-space lines?	No
CJ_merged_000088	[ { "path": "images/Dronemillion/img_008301.jpg" } ]	Is there a person wearing bright yellow clothing to the left of the pink car in the parking area beside the intersection at the lower right of the image?	Yes
CJ_merged_000089	[ { "path": "images/Dronemillion/img_008311.jpg" } ]	Given the heavy haze over the intersection, are vehicles still moving through it while the lane markings and crosswalks remain clearly visible enough to guide traffic?	Yes
CJ_merged_000090	[ { "path": "images/Dronemillion/img_008321.jpg" } ]	Is the white van near the center of the intersection positioned fully within a marked lane and clearly aligned with the painted lane arrows?	No
CJ_merged_000091	[ { "path": "images/Dronemillion/img_008331.jpg" } ]	Despite the heavy haze, are vehicles at the central intersection generally aligned within their lanes and stopping outside the pedestrian crosswalk zones?	Yes
CJ_merged_000092	[ { "path": "images/Dronemillion/img_008338.jpg" } ]	Is the green city bus both stopped entirely within its lane and aligned parallel to the lane markings without encroaching on the adjacent crosswalk?	Yes
CJ_merged_000093	[ { "path": "images/Dronemillion/img_008352.jpg" } ]	In the lower-right parking area, are all vehicles parked in marked spaces, and is there no one walking on the crosswalk to the left?	No
CJ_merged_000094	[ { "path": "images/Dronemillion/img_008364.jpg" } ]	In the parking area in front of the curved white building on the right, are some cars parked in an orderly row while others are angled or double-parked, suggesting partial parking disorder?	Yes
CJ_merged_000095	[ { "path": "images/Dronemillion/img_008373.jpg" } ]	In the parking area in front of the curved white building, are all vehicles parked neatly within marked spaces and oriented in the same direction toward the building?	No
CJ_merged_000096	[ { "path": "images/Dronemillion/img_008380.jpg" } ]	At the main intersection, are vehicles stopped behind the pedestrian crosswalk lines while the signal for their direction appears to be red or otherwise not permitting movement?	No
CJ_merged_000098	[ { "path": "images/Dronemillion/img_008401.jpg" } ]	Are there pedestrians present on the bottom crosswalk or its adjacent sidewalk while nearby vehicles are waiting close to the stop line?	No
CJ_merged_000099	[ { "path": "images/Dronemillion/img_008412.jpg" } ]	Are all vehicles in the multi-lane approach at the right edge properly aligned within single lanes and stopped behind the nearest crosswalk line?	No
CJ_merged_000100	[ { "path": "images/Dronemillion/img_008419.jpg" } ]	Are the vehicles queuing along the curved curb at the right edge mostly aligned within a single lane and oriented in the same direction, suggesting an orderly waiting queue?	Yes
CJ_merged_000101	[ { "path": "images/Dronemillion/img_008428.jpg" } ]	Is motor vehicle traffic at the intersection generally orderly, with cars staying within marked lanes and no vehicles stopped inside the central box area?	Yes
CJ_merged_000102	[ { "path": "images/Dronemillion/img_008435.jpg" } ]	Across the whole intersection, are vehicles generally aligned within marked lanes and crosswalk areas kept clear of stopped vehicles?	No
CJ_merged_000103	[ { "path": "images/Dronemillion/img_008446.jpg" } ]	Are all vehicles in this right-side queue properly aligned within their lanes and leaving a clear gap behind the red bus ahead of them?	No
CJ_merged_000104	[ { "path": "images/Dronemillion/img_008454.jpg" } ]	At this central intersection, are all vehicle lanes with straight-through arrows completely free of stopped vehicles while every pedestrian waiting area at the crosswalks is occupied by at least one person?	No
CJ_merged_000105	[ { "path": "images/Dronemillion/img_008462.jpg" } ]	On the central pedestrian street, are decorative lanterns hanging overhead while several pedestrians walk along the wet, snow-lined pathway?	Yes
CJ_merged_000107	[ { "path": "images/Dronemillion/img_008477.jpg" } ]	Is the orange lifting machine positioned on cleared pavement while its boom extends over snow-covered ground where no pedestrians are standing nearby?	Yes
CJ_merged_000110	[ { "path": "images/Dronemillion/img_008492.jpg" } ]	Is the orange boom lift positioned on the snowy ground with its boom extended over the cleared area while nearby cars remain parked outside its immediate working range?	No
CJ_merged_000112	[ { "path": "images/Dronemillion/img_008503.jpg" } ]	Is the pedestrian plaza fully cleared of snow with all maintenance workers concentrated there and no remaining snow piles on the walking paths?	No
CJ_merged_000114	[ { "path": "images/Dronemillion/img_008513.jpg" } ]	On the circular ice rink, are all skaters moving in the same clockwise direction while staying within the inner ring close to the central monument?	No
CJ_merged_000115	[ { "path": "images/Dronemillion/img_008520.jpg" } ]	Are all people on the circular ice rink wearing dark clothing and moving in the same counterclockwise direction around the center pillar?	No
CJ_merged_000117	[ { "path": "images/Dronemillion/img_008530.jpg" } ]	On the circular ice rink in the center, are all skaters moving in an organized single clockwise direction while being supervised by visible staff standing evenly around the rink’s perimeter?	No
CJ_merged_000120	[ { "path": "images/Dronemillion/img_008542.jpg" } ]	On the circular ice rink in the center, are all skaters wearing dark clothing with no one in a light-colored outfit such as white or bright red?	No
CJ_merged_000121	[ { "path": "images/Dronemillion/img_008544.jpg" } ]	Are all people on the circular skating area staying within the blue-lit rink boundary and wearing similarly bright clothing suitable for skating?	No
CJ_merged_000124	[ { "path": "images/Dronemillion/img_008545.jpg" } ]	In the circular plaza, are all people on the ice confined within the rink barrier and wearing clearly visible helmets?	No
CJ_merged_000127	[ { "path": "images/Dronemillion/img_008552.jpg" } ]	Are some of the cars along the lower road positioned such that they partially block driveways or turning space while pedestrians are walking nearby on the roadside?	Yes
CJ_merged_000128	[ { "path": "images/Dronemillion/img_008556.jpg" } ]	On the central road, are vehicles forming a congested queue while pedestrians are simultaneously walking along both sides rather than using a separated sidewalk area?	Yes
CJ_merged_000129	[ { "path": "images/Dronemillion/img_008561.jpg" } ]	Are the people in this long queue arranged in a single orderly line with clear spacing between individuals and no side-by-side clusters forming secondary lines?	No
CJ_merged_000130	[ { "path": "images/Dronemillion/img_008565.jpg" } ]	Are all cars in this lower-right courtyard row parked neatly within single marked spaces and facing the same direction toward the building?	No
CJ_merged_000131	[ { "path": "images/Dronemillion/img_008569.jpg" } ]	Does the long queue along the lit sidewalk consist of people spaced apart in an orderly single line while several staff-like figures stand separately near the bright entrance area?	No
CJ_merged_000132	[ { "path": "images/Dronemillion/img_008574.jpg" } ]	Is the crowd along the illuminated sidewalk mostly standing in a single straight line with clear spacing between individuals and without people spilling onto the adjacent roadway?	No
CJ_merged_000133	[ { "path": "images/Dronemillion/img_008583.jpg" } ]	Is the box truck near the middle-right edge parked neatly within a clearly marked parking space and oriented parallel to the adjacent cars?	No
CJ_merged_000134	[ { "path": "images/Dronemillion/img_008587.jpg" } ]	On the central horizontal road, are all visible vehicles parked neatly within marked parking bays and aligned parallel to the road edges?	No
CJ_merged_000136	[ { "path": "images/Dronemillion/img_008598.jpg" } ]	Within this residential block, are most parked cars positioned in marked courtyard or roadside spaces while the internal driveways remain largely clear for vehicle passage?	Yes
CJ_merged_000137	[ { "path": "images/Dronemillion/img_008606.jpg" } ]	Across the whole scene, are all vehicles parked or stopped only in clearly designated parking areas, with no cars standing along the narrow inner roads?	No
CJ_merged_000138	[ { "path": "images/Dronemillion/img_008613.jpg" } ]	In the central vertical parking area, are most parking spaces occupied by cars while the driveway between the rows is kept clear for vehicle passage?	Yes
CJ_merged_000140	[ { "path": "images/Dronemillion/img_008624.jpg" } ]	Within this lit compound, are multiple workers clearly visible near the bright floodlights and is there heavy machinery operating under those lights?	No
CJ_merged_000149	[ { "path": "images/Dronemillion/img_008643.jpg" } ]	Is the large box truck beside the warehouse positioned within a marked parking bay and aligned roughly parallel to the painted bay lines?	No
CJ_merged_000151	[ { "path": "images/Dronemillion/img_008653.jpg" } ]	Are the vehicles in the long queue on the curved middle road positioned in clearly marked lanes with visible lane dividers and a traffic signal controlling their wait?	No
CJ_merged_000152	[ { "path": "images/Dronemillion/img_008658.jpg" } ]	Is the yellow construction vehicle positioned in the yard both oriented parallel to the nearby parked cars and surrounded by other vehicles indicating it is currently parked rather than operating?	No
CJ_merged_000154	[ { "path": "images/Dronemillion/img_008663.jpg" } ]	Are the cars parked along the center road positioned in an orderly single line on one side of the road, leaving the opposite side generally clear for vehicle passage?	Yes
CJ_merged_000155	[ { "path": "images/Dronemillion/img_008674.jpg" } ]	Is the bridge wide enough to carry two clearly marked lanes in each direction with continuous lane markings and visible pedestrian sidewalks on both sides?	No
CJ_merged_000157	[ { "path": "images/Dronemillion/img_008682.jpg" } ]	Is at least one vehicle on this main road driving in each direction, with the vehicles positioned in the appropriate lanes separated by the central markings?	Yes
CJ_merged_000159	[ { "path": "images/Dronemillion/img_008690.jpg" } ]	Are all vehicles traveling on the main road properly using only one lane each, with no cars straddling the lane markings or partially stopped on the roadside?	No
CJ_merged_000160	[ { "path": "images/Dronemillion/img_008693.jpg" } ]	In the open yard between the two blue-roofed buildings, are there both curved internal pathways and at least two distinct circular features visible on the ground?	Yes
CJ_merged_000161	[ { "path": "images/Dronemillion/img_008701.jpg" } ]	In the large paved parking lot in the upper left, are cars parked only within marked bays while several rows of spaces remain entirely empty, indicating low occupancy but orderly parking?	Yes
CJ_merged_000162	[ { "path": "images/Dronemillion/img_008705.jpg" } ]	Does the whole scene show a clearly maintained separation between the factory buildings and the frozen waterway, with a continuous secure fence and no apparent breaches or gaps?	No
CJ_merged_000163	[ { "path": "images/Dronemillion/img_008707.jpg" } ]	Does the long industrial building at the top left have both multiple rooftop ventilation units and several vehicles parked directly along its front side?	Yes
CJ_merged_000165	[ { "path": "images/Dronemillion/img_008723.jpg" } ]	Across the whole scene, are vehicles generally parked in an orderly manner along designated edges of roads and yards without visibly blocking building entrances or internal driveways?	Yes
CJ_merged_000166	[ { "path": "images/Dronemillion/img_008727.jpg" } ]	In this open yard between the two large buildings, are multiple cars parked in an orderly manner while the ground also shows scattered snow patches and construction materials or debris?	Yes
CJ_merged_000172	[ { "path": "images/Dronemillion/img_008748.jpg" } ]	In the parking area beside the main road, are there multiple cars parked in an organized, front-in orientation rather than randomly scattered or angled?	Yes
CJ_merged_000173	[ { "path": "images/Dronemillion/img_008752.jpg" } ]	In the open paved area in front of the tall beige high‑rise, are pedestrians present while several cars are parked in an orderly manner around its edges?	Yes
CJ_merged_000174	[ { "path": "images/Dronemillion/img_008758.jpg" } ]	Are the cars parked along the edge of the open yard arranged in a single orderly row, all facing toward the yard rather than toward the street?	Yes
CJ_merged_000175	[ { "path": "images/Dronemillion/img_008763.jpg" } ]	Are several passenger cars parked in marked spaces and aligned consistently within this parking lot beside the tall red-roofed building?	No
CJ_merged_000176	[ { "path": "images/Dronemillion/img_008766.jpg" } ]	Are all vehicles parked along this narrow road uniformly aligned within a clearly marked parking area without any vehicle encroaching onto the roadway?	No
CJ_merged_000178	[ { "path": "images/Dronemillion/img_008778.jpg" } ]	Within the open dirt yard beside the parked vehicles, is there clear evidence of both heavy vehicle tire tracks and scattered material piles indicating ongoing industrial or construction activity?	Yes
CJ_merged_000179	[ { "path": "images/Dronemillion/img_008782.jpg" } ]	On the multi-lane road in the middle, are moving vehicles present while the lanes closest to the upper roadside are mostly occupied by a continuous line of parked cars?	Yes
CJ_merged_000180	[ { "path": "images/Dronemillion/img_008790.jpg" } ]	In the main parking lot beside the long rectangular building, are all parked cars aligned neatly within marked spaces without any vehicle encroaching into the drive lanes or snow-covered areas?	No
CJ_merged_000181	[ { "path": "images/Dronemillion/img_008794.jpg" } ]	On this road segment, are both traffic lanes fully occupied by vehicles moving in opposite directions with no long empty gaps between cars?	No
CJ_merged_000182	[ { "path": "images/Dronemillion/img_008797.jpg" } ]	Does the central farmland show regularly spaced, curved planting rows with narrow irrigation channels between the row blocks and patches of remaining snow along some of the boundaries?	Yes
CJ_merged_000183	[ { "path": "images/Dronemillion/img_008801.jpg" } ]	On this middle road, are all vehicles parked neatly within marked spaces and facing the same direction along the curb?	No
CJ_merged_000185	[ { "path": "images/Dronemillion/img_008812.jpg" } ]	Does this industrial building show signs of being in active use, with vehicles parked immediately beside it and its roof areas largely intact rather than collapsed?	Yes
CJ_merged_000186	[ { "path": "images/Dronemillion/img_008818.jpg" } ]	Do the cars parked along the road in front of the pagoda-like building appear orderly in a single row and positioned fully within the paved roadside area rather than on the adjacent paths or grass?	No
CJ_merged_000187	[ { "path": "images/Dronemillion/img_008824.jpg" } ]	Within this central rectangular plot, are most planting rows fully covered by snow with distinct untouched white bands running continuously from one side of the plot to the other?	No

End of preview.

Drone-PUR Benchmark

Paper | Project

This package provides the test split of Drone-PUR, covering perception, understanding, and reasoning in UAV scenes: 11 tasks, 20,051 test questions, and 22,248 image files. The ten choice-based tasks, including binary CJ, contain 18,431 questions; OD uses open-ended multi-box prediction.

Images come from eight established datasets: VisDrone, DroneVehicle, DroneCrowd, MDMT, MDOT, DroneRGBT, AnimalDrone, and DroneBird. Newly collected images are stored under images/Dronemillion/, corresponding to the paper's "New Images". The image counts in this package refer to the test split, not the full dataset. GPS location fields have been removed from the released Dronemillion images; image pixels, dimensions, orientation, and embedded frames are preserved.

Tasks

Dimension	Task	Questions
Perception	Object Detection (OD)	1,620
Perception	Object Recognition (OR)	1,785
Perception	Object Counting (OC)	1,761
Perception	Multi-Image Grounding (MIG)	990
Understanding	Conditional Judgment (CJ)	2,512
Understanding	Visual Grounding (VG)	1,361
Understanding	Position Spatial Judgment (PSJ)	2,719
Reasoning	MMR-C: joint counting	1,708
Reasoning	MMR-G: joint grounding	1,211
Reasoning	MMR-R: multimodal reasoning	615
Reasoning	MVR-G: cross-view grounding	1,241
Reasoning	MVR-R: multi-view reasoning	424
Reasoning	Causal Reasoning (CR)	1,031
Reasoning	Situational Inference (SI)	1,073

MMR's three subtasks form one task; MVR's two subtasks form one task. Paper names MMR-C, MMR-G, MVR-G correspond to file/loader identifiers MMR-OC, MMR-VG, MVR-VG, respectively. Other identifiers are unchanged.

PSJ includes vehicle heading (VH, 359), pedestrian heading (PH, 341), turning direction (TD, 677), horizontal relative position (RP, 1,173), and height comparison (HC, 169). VH/PH use image-relative headings; TD/RP use object-relative directions. HC compares elevation, not object height. The stored label 3. Travel Direction refers to TD.

Files and Annotations

Drone-PUR-Benchmark/
├── annotations/          # Task folders; MMR/MVR have separate subtask JSON files
├── images/<source>/     # Images referenced by the test questions
├── metadata/
│   ├── image_manifest.json
│   └── template_config.json
├── tools/
│   ├── benchmark_io.py
│   └── prepare_templates.py
├── requirements.txt
├── LICENSE
└── README.md

Each JSON file has dataset, split, task, subtask, total_samples, and results. Single-task files use annotations/<TASK>/test.json; MMR uses test_oc.json, test_vg.json, test_reasoning.json, and MVR uses the latter two.

Each sample has question_id, question, and an ordered images list whose entries contain only a root-relative path. Choice samples also have options and correct_answer. CJ stores Yes/No directly, without an option dictionary.

ground_truth stores open-ended answers: count, box, or boxes, with a type and, for localization, a 1-based image_index.
MIG and MVR-G have template.box identifying the reference target in Image 1. MIG's query_frame identifies the answer frame.
OR has a target bbox. PSJ has sampled_question_type and applicable object boxes. HC's high_box/low_box are answer-related references, not model inputs.

Boxes use normalized [x1, y1, x2, y2] coordinates in the indicated image, with the origin at the upper-left. MVR-G absent targets use type: "absent" and the not exist choice, not a zero-area box. Do not pass answers to models; omit candidate options for open-ended prediction.

image_manifest.json provides image paths, sources, dimensions, sizes, and checksums. template_config.json provides shared reference-box drawing settings.

Download

Download the repository snapshot while keeping its directory layout:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Henard224/Drone-PUR-Benchmark",
    repo_type="dataset",
    local_dir="./Drone-PUR-Benchmark",
)

The 14 Hugging Face configurations correspond to the annotation files, each with a test split. They expose annotation rows; image paths remain relative to the downloaded repository. The original JSON files retain their headers and results arrays.

Image Order and Templates

Task	Input order
MIG	Reference image first, then three later frames; answer in `query_frame`
MMR	RGB first, TIR second; MMR-G boxes refer to RGB
MVR	View 1 then View 2; MVR-G matches the reference target into View 2

Before evaluating MIG or MVR-G, generate the red-box reference images:

python -m pip install -r requirements.txt
python tools/prepare_templates.py

Python 3.10+ is required. Generated images are saved under prepared_images/; source images are not overwritten. Only Image 1 receives the reference box. The drawing utility reads metadata/template_config.json and reuses matching outputs on subsequent runs. Keep the original image order and do not draw twice.

Usage

Run template preparation first for MIG and MVR-G.

import sys
from pathlib import Path

root = Path('/path/to/Drone-PUR-Benchmark')
sys.path.insert(0, str(root / 'tools'))
from benchmark_io import load_annotations, model_image_paths, get_ground_truth

sample = load_annotations(root, 'MVR-VG')[0]  # MVR-G in the paper
image_paths = model_image_paths(root, sample)
question = sample['question']
options = sample['options']
reference = get_ground_truth(sample)  # For evaluation only.

The reader supplies task/subtask context from the JSON header. It selects the prepared reference image where required and leaves other inputs unchanged.

The paper reports accuracy for choice-based tasks, Acc@0.5 for grounding, MAE/MRA for counting, and Micro-F1@0.5 for OD. Grounding correctness requires IoU > 0.5, or a correct not exist response for absent targets. Aggregate PSJ, MMR, and MVR by sample count; overall choice accuracy is sample-weighted across the 18,431 choice/Yes-No questions.

This dataset repository contains no inference or scoring programs. The included tools only read annotations and prepare reference images. Evaluation code belongs to the separate GitHub code release; see the project page for code availability.

License

Drone-PUR Benchmark is released under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). Sharing and adaptation are permitted for non-commercial purposes with attribution; adaptations must use the same or a compatible license. See LICENSE for the complete terms. Previously released source images retain their original rights and applicable terms; no third-party rights are overridden by this notice. Please cite the linked paper when using Drone-PUR.

Downloads last month: 187

Paper for Henard224/Drone-PUR-Benchmark

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

Paper • 2610.40801 • Published 10 days ago