NF-UAVTrack / README.md
windowflower's picture
Update dataset card
210908a verified
|
Raw History Blame Contribute Delete
8 kB
---
license: cc-by-4.0
task_categories:
- object-detection
tags:
- uav
- drone
- anti-uav
- night
- fog
- low-light
- small-object-tracking
- single-object-tracking
- uav-tracking
- yolo
pretty_name: NF-UAVTrack
size_categories:
- 10K<n<100K
---
# NF-UAVTrack
**NF-UAVTrack** (Night-and-Fog UAV Tracking) is a visible-light **anti-UAV** dataset
for **single-object tracking and detection of small UAVs**. All videos were filmed
from the ground with smartphones at night in fog, and the flying UAV target is
annotated frame by frame in YOLO format.
- **25 videos**, **12,127 frames**, about **1.2 GB**
- **Four scene types** (open grassland, sports field, park, between buildings) and **two recording periods** (April 2025 and April 2026)
- **10,936** frames with exactly one target box and **1,191** target-absent frames
(empty label files)
- Resolutions: **3840×2160** (video01–video09) and **1920×1080** (video10–video25)
- Small targets: the mean box covers **2.67 % × 3.96 %** of the image width × height,
and the smallest box covers **0.75 % × 1.06 %**
- The target leaves the view or becomes invisible in **20 of the 25 videos**, in
**108** separate absence intervals
## Collection
All videos were recorded by the authors from the ground with two smartphones
(Xiaomi 12S and iPhone 15 Pro) in four different areas of the Future Campus of Soochow
University, Suzhou, Jiangsu, China, one per scene type (open grassland, sports field,
park and between buildings), in two recording periods:
April 2025 and April 2026. Recording took place in foggy
conditions between 23:00 and 01:00 the following day (China Standard Time, UTC+8).
The targets are a YH F450 and a DJI Mavic 4 in flight.
Frames were extracted from the videos, and the target was annotated manually, frame
by frame, with an axis-aligned bounding box. Two annotators labelled separate sets of
videos and reviewed each other's annotations. Each sequence `video01` … `video25` is
one continuous recording of one target, with the original frame order preserved.
## Dataset structure
```
NF-UAVTrack/
├── README.md
├── LICENSE
├── metadata.csv
├── images/
│ ├── video01/
│ │ ├── frame-000001.jpg
│ │ └── ...
│ └── ... # video01 … video25
└── labels/
├── video01/
│ ├── frame-000001.txt
│ └── ... # same basename as the corresponding image
└── ...
```
Every image has exactly one label file with the same basename, and every label
file has an image. There are no orphan files.
`metadata.csv` lists every frame:
| column | meaning |
|---|---|
| `video` | sequence id, e.g. `video01` |
| `frame` | frame id, e.g. `frame-000001` |
| `image_path` | image path relative to the repository root |
| `label_path` | label path relative to the repository root |
| `has_target` | `1` if the frame has a target box, `0` if the frame is target-absent |
## Label format
Labels use the standard **YOLO detection** format: one text file per frame, with
coordinates normalised to `[0, 1]` by the image width and height.
```
class_id x_center y_center width height
```
Example (`labels/video01/frame-000001.txt`):
```
0 0.444961 0.817969 0.031459 0.040373
```
- There is a **single class**, `class_id = 0`, the UAV target.
- Each annotated frame has **exactly one box**.
- **Empty label files are intentional.** A 0-byte `.txt` file marks a
**target-absent** frame: the target has left the field of view or is not visible
in the fog. Keep these files, because trackers are evaluated on whether they
handle the target's disappearance and reappearance.
## Official split
The accompanying paper uses a fixed split by video:
- **Train (14 videos):** video01, video03, video04, video05, video06, video07,
video10, video11, video12, video15, video17, video18, video21, video22
- **Test (11 videos):** video02, video08, video09, video13, video14, video16,
video19, video20, video23, video24, video25
Split files: <https://github.com/windowFlowers/NF-UAVTrack/tree/main/splits>
Split by video, not by frame. Neighbouring frames of one video are nearly
identical, so a random frame-level split leaks test data into training.
## Statistics
| Video | Resolution | Frames | With target | Target-absent |
|---|---|---:|---:|---:|
| video01 | 3840×2160 | 583 | 333 | 250 |
| video02 | 3840×2160 | 775 | 775 | 0 |
| video03 | 3840×2160 | 2488 | 2488 | 0 |
| video04 | 3840×2160 | 443 | 401 | 42 |
| video05 | 3840×2160 | 296 | 233 | 63 |
| video06 | 3840×2160 | 335 | 104 | 231 |
| video07 | 3840×2160 | 566 | 557 | 9 |
| video08 | 3840×2160 | 157 | 140 | 17 |
| video09 | 3840×2160 | 520 | 477 | 43 |
| video10 | 1920×1080 | 300 | 299 | 1 |
| video11 | 1920×1080 | 300 | 299 | 1 |
| video12 | 1920×1080 | 300 | 288 | 12 |
| video13 | 1920×1080 | 300 | 300 | 0 |
| video14 | 1920×1080 | 499 | 374 | 125 |
| video15 | 1920×1080 | 257 | 257 | 0 |
| video16 | 1920×1080 | 375 | 234 | 141 |
| video17 | 1920×1080 | 751 | 595 | 156 |
| video18 | 1920×1080 | 193 | 186 | 7 |
| video19 | 1920×1080 | 168 | 138 | 30 |
| video20 | 1920×1080 | 814 | 795 | 19 |
| video21 | 1920×1080 | 319 | 310 | 9 |
| video22 | 1920×1080 | 295 | 295 | 0 |
| video23 | 1920×1080 | 149 | 126 | 23 |
| video24 | 1920×1080 | 261 | 254 | 7 |
| video25 | 1920×1080 | 683 | 678 | 5 |
| **Total** | | **12,127** | **10,936** | **1,191** |
Box size (normalised by image width / height):
| | min | mean | max |
|---|---:|---:|---:|
| width | 0.0075 | 0.0267 | 0.1139 |
| height | 0.0106 | 0.0396 | 0.1557 |
## Intended use
- Evaluating and training **single-object trackers** for anti-UAV surveillance at
night and in fog, including target disappearance and reappearance.
- Training and evaluating **small-UAV detectors** on ground-to-air images taken in
low light.
## Limitations
- One class and one target per video; not suited to multi-class or multi-object
benchmarks.
- Night and fog only; daytime and clear-weather scenes are not covered.
- Two resolutions (3840×2160 and 1920×1080) are mixed. Resize consistently or report
results per resolution group.
- Target-absent frames are unevenly distributed (0 % to 69 % of a video), so
per-video results can vary widely.
## Loading
```python
from datasets import load_dataset
# frame list with paths and has_target flags
ds = load_dataset("windowflower/NF-UAVTrack", data_files="metadata.csv", split="train")
print(ds[0])
# download a single image
from huggingface_hub import hf_hub_download
path = hf_hub_download("windowflower/NF-UAVTrack",
"images/video01/frame-000001.jpg", repo_type="dataset")
```
Note that `split="train"` above only names the loaded table; it is not the official
training split. Use the split files listed above.
To download everything:
```bash
git lfs install
git clone https://huggingface.co/datasets/windowflower/NF-UAVTrack
```
## License
Released under the **Creative Commons Attribution 4.0 International (CC BY 4.0)**
licence. See [LICENSE](LICENSE). You may share and adapt the data, including
commercially, as long as you give appropriate credit and indicate any changes.
## Citation
If you use this dataset, please cite:
```bibtex
@misc{cai_nfuavtrack_2026,
title = {NF-UAVTrack: A Night-and-Fog Visible-Light Anti-UAV Tracking Dataset},
author = {Cai, Zhexuan},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/windowflower/NF-UAVTrack}},
note = {Source repository: \url{https://github.com/windowFlowers/NF-UAVTrack}. Licensed under CC BY 4.0}
}
```
The accompanying paper is under review; its reference will be added here once it is
published.
The GitHub repository <https://github.com/windowFlowers/NF-UAVTrack> hosts the
official split files and the annotation documentation.
## Contact
For questions, corrections or annotation issues, please open a discussion on this
repository.