windowflower commited on
Commit
01dd7fe
·
verified ·
1 Parent(s): da9b72c

Replace dataset card with revised README

Browse files
Files changed (1) hide show
  1. README.md +116 -106
README.md CHANGED
@@ -2,15 +2,17 @@
2
  license: cc-by-4.0
3
  task_categories:
4
  - object-detection
5
- - video-classification
6
  tags:
7
  - uav
8
- - aerial-imagery
 
 
 
 
9
  - small-object-tracking
10
  - single-object-tracking
11
- - yolo
12
- - drone
13
  - uav-tracking
 
14
  pretty_name: NF-UAVTrack
15
  size_categories:
16
  - 10K<n<100K
@@ -18,16 +20,32 @@ size_categories:
18
 
19
  # NF-UAVTrack
20
 
21
- A frame-level UAV (Unmanned Aerial Vehicle) aerial tracking dataset for **single-object
22
- detection and tracking of small targets**, recorded from an onboard drone camera and
 
23
  annotated frame by frame in YOLO format.
24
 
25
- - **25 video sequences**, **12,127 frames**, **1,226 MiB**
26
- - **10,936** positively annotated frames (exactly one target box each) and
27
- **1,191** explicitly empty (negative) frames
28
- - Resolutions: **3840×2160** (9 sequences) and **1920×1080** (16 sequences)
29
- - Targets are small: the average ground-truth box covers only
30
- **2.67 % × 3.96 %** of the image, with boxes as small as **0.75 % × 1.06 %**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
31
 
32
  ## Dataset structure
33
 
@@ -40,18 +58,16 @@ NF-UAVTrack/
40
  │ ├── video01/
41
  │ │ ├── frame-000001.jpg
42
  │ │ └── ...
43
- │ ├── video02/
44
  │ └── ... # video01 … video25
45
  └── labels/
46
  ├── video01/
47
  │ ├── frame-000001.txt
48
  │ └── ... # same basename as the corresponding image
49
- ├── video02/
50
  └── ...
51
  ```
52
 
53
- Every image has exactly one label file with a matching basename, and vice versa.
54
- There are no orphan images or orphan labels.
55
 
56
  `metadata.csv` lists every frame:
57
 
@@ -59,14 +75,14 @@ There are no orphan images or orphan labels.
59
  |---|---|
60
  | `video` | sequence id, e.g. `video01` |
61
  | `frame` | frame id, e.g. `frame-000001` |
62
- | `image_path` | path relative to the repository root |
63
- | `label_path` | path relative to the repository root |
64
- | `has_target` | `1` if the frame contains an annotated target, `0` if it is an empty (negative) frame |
65
 
66
  ## Label format
67
 
68
- Labels follow the standard **YOLO detection** convention, one text file per frame,
69
- one line per object, coordinates normalised to `[0, 1]`:
70
 
71
  ```
72
  class_id x_center y_center width height
@@ -78,106 +94,101 @@ Example (`labels/video01/frame-000001.txt`):
78
  0 0.444961 0.817969 0.031459 0.040373
79
  ```
80
 
81
- Notes on this release:
 
 
 
 
 
82
 
83
- - A **single class**, `class_id = 0`, corresponding to the tracked UAV target.
84
- - Exactly **one box per positive frame** — this is a single-target tracking dataset.
85
- - **Empty label files are intentional.** A 0-byte `.txt` file means the target is not
86
- visible in that frame (occlusion, out of frame, or too small to annotate reliably).
87
- Do not discard these files: they carry the negative-sample signal that lets a
88
- detector learn to suppress spurious detections.
89
 
90
- ## Statistics
91
 
92
- | Video | Frames | With target | Empty |
93
- |---|---|---|---|
94
- | video01 | 583 | 333 | 250 |
95
- | video02 | 775 | 775 | 0 |
96
- | video03 | 2488 | 2488 | 0 |
97
- | video04 | 443 | 401 | 42 |
98
- | video05 | 296 | 233 | 63 |
99
- | video06 | 335 | 104 | 231 |
100
- | video07 | 566 | 557 | 9 |
101
- | video08 | 157 | 140 | 17 |
102
- | video09 | 520 | 477 | 43 |
103
- | video10 | 300 | 299 | 1 |
104
- | video11 | 300 | 299 | 1 |
105
- | video12 | 300 | 288 | 12 |
106
- | video13 | 300 | 300 | 0 |
107
- | video14 | 499 | 374 | 125 |
108
- | video15 | 257 | 257 | 0 |
109
- | video16 | 375 | 234 | 141 |
110
- | video17 | 751 | 595 | 156 |
111
- | video18 | 193 | 186 | 7 |
112
- | video19 | 168 | 138 | 30 |
113
- | video20 | 814 | 795 | 19 |
114
- | video21 | 319 | 310 | 9 |
115
- | video22 | 295 | 295 | 0 |
116
- | video23 | 149 | 126 | 23 |
117
- | video24 | 261 | 254 | 7 |
118
- | video25 | 683 | 678 | 5 |
119
- | **Total** | **12,127** | **10,936** | **1,191** |
120
-
121
- Box size distribution (normalised, relative to image width/height):
122
-
123
- | — | min | mean | max |
124
- |---|---|---|---|
125
- | width | 0.0075 | 0.0267 | 0.1139 |
126
- | height | 0.0106 | 0.0396 | 0.1557 |
127
 
128
- Resolution split:
129
 
130
- | Sequences | Resolution | Frames |
131
- |---|---|---|
132
- | video01 – video09 | 3840 × 2160 | 6,163 |
133
- | video10 – video25 | 1920 × 1080 | 5,964 |
134
-
135
- ## Collection
136
 
137
- <!-- TODO before publishing: fill in the recording details you know for certain —
138
- drone model, camera/lens, flight altitude range, typical ground-sample distance,
139
- capture frame rate, dates and locations of the flights, and the annotation tool
140
- (e.g. LabelImg / CVAT / custom). Reviewers of a dataset paper will ask for these. -->
141
 
142
- All footage was recorded by the authors with a drone camera. Frames were extracted
143
- from the raw flight recordings and the tracked target was annotated manually, frame
144
- by frame, with an axis-aligned bounding box. Sequences were then split into
145
- `video01` … `video25`, each corresponding to one continuous flight segment / target.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
146
 
147
  ## Intended use
148
 
149
- - Training and evaluating **small-object detectors** on aerial imagery.
150
- - Training and evaluating **single-object trackers** (SOT) where the target may leave
151
- the field of view, in which case the frame is labelled as empty.
152
- - Benchmarking robustness to very small objects, since the mean target occupies less
153
- than 3 % of the image width.
154
 
155
  ## Limitations
156
 
157
- - Single class only; not suitable for multi-class detection benchmarks.
158
- - One target per sequence — the dataset is not designed for crowded multi-object
159
- scenes.
160
- - The 4K and 1080p sequences come from different capture configurations; report
161
- results per resolution group or resample before mixing them.
162
- - Empty frames are scene-dependent and highly unbalanced across sequences (from 0 %
163
- to 69 % of a sequence), so per-sequence metrics may vary a lot.
164
 
165
  ## Loading
166
 
167
  ```python
168
  from datasets import load_dataset
169
 
170
- # frame list + labels as metadata
171
  ds = load_dataset("windowflower/NF-UAVTrack", data_files="metadata.csv", split="train")
172
  print(ds[0])
173
 
174
- # images are read directly from the repo
175
  from huggingface_hub import hf_hub_download
176
  path = hf_hub_download("windowflower/NF-UAVTrack",
177
  "images/video01/frame-000001.jpg", repo_type="dataset")
178
  ```
179
 
180
- Or clone everything:
 
 
 
181
 
182
  ```bash
183
  git lfs install
@@ -186,17 +197,17 @@ git clone https://huggingface.co/datasets/windowflower/NF-UAVTrack
186
 
187
  ## License
188
 
189
- Released under **Creative Commons Attribution 4.0 International (CC BY 4.0)**.
190
- See [LICENSE](LICENSE). You may share and adapt the data, including commercially,
191
- as long as you give appropriate credit and indicate changes.
192
 
193
  ## Citation
194
 
195
- If you use this dataset, please cite the dataset and the accompanying paper:
196
 
197
  ```bibtex
198
  @misc{cai_nfuavtrack_2026,
199
- title = {NF-UAVTrack: A UAV Aerial Small-Object Tracking Dataset},
200
  author = {Cai, Zhexuan},
201
  year = {2026},
202
  howpublished = {\url{https://huggingface.co/datasets/windowflower/NF-UAVTrack}},
@@ -204,14 +215,13 @@ If you use this dataset, please cite the dataset and the accompanying paper:
204
  }
205
  ```
206
 
207
- The GitHub repository at
208
- <https://github.com/windowFlowers/NF-UAVTrack> mirrors this dataset and hosts the
209
- official train/test split, per-sequence statistics, and the annotation format
210
- documentation.
211
 
212
- <!-- TODO before publishing: add the paper reference (authors, title, venue, year)
213
- once it is public, and keep the dataset URL as the stable identifier. -->
214
 
215
  ## Contact
216
 
217
- Questions, corrections, or annotation issues: open a discussion on this repository.
 
 
2
  license: cc-by-4.0
3
  task_categories:
4
  - object-detection
 
5
  tags:
6
  - uav
7
+ - drone
8
+ - anti-uav
9
+ - night
10
+ - fog
11
+ - low-light
12
  - small-object-tracking
13
  - single-object-tracking
 
 
14
  - uav-tracking
15
+ - yolo
16
  pretty_name: NF-UAVTrack
17
  size_categories:
18
  - 10K<n<100K
 
20
 
21
  # NF-UAVTrack
22
 
23
+ **NF-UAVTrack** (Night-and-Fog UAV Tracking) is a visible-light **anti-UAV** dataset
24
+ for **single-object tracking and detection of small UAVs**. All videos were filmed
25
+ from the ground with smartphones at night in fog, and the flying UAV target is
26
  annotated frame by frame in YOLO format.
27
 
28
+ - **25 videos**, **12,127 frames**, about **1.2 GB**
29
+ - **10,936** frames with exactly one target box and **1,191** target-absent frames
30
+ (empty label files)
31
+ - Resolutions: **3840×2160** (video01–video09) and **1920×1080** (video10–video25)
32
+ - Small targets: the mean box covers **2.67 % × 3.96 %** of the image width × height,
33
+ and the smallest box covers **0.75 % × 1.06 %**
34
+ - The target leaves the view or becomes invisible in **20 of the 25 videos**, in
35
+ **108** separate absence intervals
36
+
37
+ ## Collection
38
+
39
+ All videos were recorded by the authors from the ground with two smartphones
40
+ (Xiaomi 12S and iPhone 15 Pro) at the Future Campus of Soochow University, Suzhou,
41
+ Jiangsu, China, in April 2025 and April 2026. Recording took place in foggy
42
+ conditions between 23:00 and 01:00 the following day (China Standard Time, UTC+8).
43
+ The targets are a YH F450 and a DJI Mavic 4 in flight.
44
+
45
+ Frames were extracted from the videos, and the target was annotated manually, frame
46
+ by frame, with an axis-aligned bounding box. Two annotators labelled separate sets of
47
+ videos and reviewed each other's annotations. Each sequence `video01` … `video25` is
48
+ one continuous recording of one target, with the original frame order preserved.
49
 
50
  ## Dataset structure
51
 
 
58
  │ ├── video01/
59
  │ │ ├── frame-000001.jpg
60
  │ │ └── ...
 
61
  │ └── ... # video01 … video25
62
  └── labels/
63
  ├── video01/
64
  │ ├── frame-000001.txt
65
  │ └── ... # same basename as the corresponding image
 
66
  └── ...
67
  ```
68
 
69
+ Every image has exactly one label file with the same basename, and every label
70
+ file has an image. There are no orphan files.
71
 
72
  `metadata.csv` lists every frame:
73
 
 
75
  |---|---|
76
  | `video` | sequence id, e.g. `video01` |
77
  | `frame` | frame id, e.g. `frame-000001` |
78
+ | `image_path` | image path relative to the repository root |
79
+ | `label_path` | label path relative to the repository root |
80
+ | `has_target` | `1` if the frame has a target box, `0` if the frame is target-absent |
81
 
82
  ## Label format
83
 
84
+ Labels use the standard **YOLO detection** format: one text file per frame, with
85
+ coordinates normalised to `[0, 1]` by the image width and height.
86
 
87
  ```
88
  class_id x_center y_center width height
 
94
  0 0.444961 0.817969 0.031459 0.040373
95
  ```
96
 
97
+ - There is a **single class**, `class_id = 0`, the UAV target.
98
+ - Each annotated frame has **exactly one box**.
99
+ - **Empty label files are intentional.** A 0-byte `.txt` file marks a
100
+ **target-absent** frame: the target has left the field of view or is not visible
101
+ in the fog. Keep these files, because trackers are evaluated on whether they
102
+ handle the target's disappearance and reappearance.
103
 
104
+ ## Official split
 
 
 
 
 
105
 
106
+ The accompanying paper uses a fixed split by video:
107
 
108
+ - **Train (14 videos):** video01, video03, video04, video05, video06, video07,
109
+ video10, video11, video12, video15, video17, video18, video21, video22
110
+ - **Test (11 videos):** video02, video08, video09, video13, video14, video16,
111
+ video19, video20, video23, video24, video25
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
112
 
113
+ Split files: <https://github.com/windowFlowers/NF-UAVTrack/tree/main/splits>
114
 
115
+ Split by video, not by frame. Neighbouring frames of one video are nearly
116
+ identical, so a random frame-level split leaks test data into training.
 
 
 
 
117
 
118
+ ## Statistics
 
 
 
119
 
120
+ | Video | Resolution | Frames | With target | Target-absent |
121
+ |---|---|---:|---:|---:|
122
+ | video01 | 3840×2160 | 583 | 333 | 250 |
123
+ | video02 | 3840×2160 | 775 | 775 | 0 |
124
+ | video03 | 3840×2160 | 2488 | 2488 | 0 |
125
+ | video04 | 3840×2160 | 443 | 401 | 42 |
126
+ | video05 | 3840×2160 | 296 | 233 | 63 |
127
+ | video06 | 3840×2160 | 335 | 104 | 231 |
128
+ | video07 | 3840×2160 | 566 | 557 | 9 |
129
+ | video08 | 3840×2160 | 157 | 140 | 17 |
130
+ | video09 | 3840×2160 | 520 | 477 | 43 |
131
+ | video10 | 1920×1080 | 300 | 299 | 1 |
132
+ | video11 | 1920×1080 | 300 | 299 | 1 |
133
+ | video12 | 1920×1080 | 300 | 288 | 12 |
134
+ | video13 | 1920×1080 | 300 | 300 | 0 |
135
+ | video14 | 1920×1080 | 499 | 374 | 125 |
136
+ | video15 | 1920×1080 | 257 | 257 | 0 |
137
+ | video16 | 1920×1080 | 375 | 234 | 141 |
138
+ | video17 | 1920×1080 | 751 | 595 | 156 |
139
+ | video18 | 1920×1080 | 193 | 186 | 7 |
140
+ | video19 | 1920×1080 | 168 | 138 | 30 |
141
+ | video20 | 1920×1080 | 814 | 795 | 19 |
142
+ | video21 | 1920×1080 | 319 | 310 | 9 |
143
+ | video22 | 1920×1080 | 295 | 295 | 0 |
144
+ | video23 | 1920×1080 | 149 | 126 | 23 |
145
+ | video24 | 1920×1080 | 261 | 254 | 7 |
146
+ | video25 | 1920×1080 | 683 | 678 | 5 |
147
+ | **Total** | | **12,127** | **10,936** | **1,191** |
148
+
149
+ Box size (normalised by image width / height):
150
+
151
+ | | min | mean | max |
152
+ |---|---:|---:|---:|
153
+ | width | 0.0075 | 0.0267 | 0.1139 |
154
+ | height | 0.0106 | 0.0396 | 0.1557 |
155
 
156
  ## Intended use
157
 
158
+ - Evaluating and training **single-object trackers** for anti-UAV surveillance at
159
+ night and in fog, including target disappearance and reappearance.
160
+ - Training and evaluating **small-UAV detectors** on ground-to-air images taken in
161
+ low light.
 
162
 
163
  ## Limitations
164
 
165
+ - One class and one target per video; not suited to multi-class or multi-object
166
+ benchmarks.
167
+ - Night and fog only; daytime and clear-weather scenes are not covered.
168
+ - Two resolutions (3840×2160 and 1920×1080) are mixed. Resize consistently or report
169
+ results per resolution group.
170
+ - Target-absent frames are unevenly distributed (0 % to 69 % of a video), so
171
+ per-video results can vary widely.
172
 
173
  ## Loading
174
 
175
  ```python
176
  from datasets import load_dataset
177
 
178
+ # frame list with paths and has_target flags
179
  ds = load_dataset("windowflower/NF-UAVTrack", data_files="metadata.csv", split="train")
180
  print(ds[0])
181
 
182
+ # download a single image
183
  from huggingface_hub import hf_hub_download
184
  path = hf_hub_download("windowflower/NF-UAVTrack",
185
  "images/video01/frame-000001.jpg", repo_type="dataset")
186
  ```
187
 
188
+ Note that `split="train"` above only names the loaded table; it is not the official
189
+ training split. Use the split files listed above.
190
+
191
+ To download everything:
192
 
193
  ```bash
194
  git lfs install
 
197
 
198
  ## License
199
 
200
+ Released under the **Creative Commons Attribution 4.0 International (CC BY 4.0)**
201
+ licence. See [LICENSE](LICENSE). You may share and adapt the data, including
202
+ commercially, as long as you give appropriate credit and indicate any changes.
203
 
204
  ## Citation
205
 
206
+ If you use this dataset, please cite:
207
 
208
  ```bibtex
209
  @misc{cai_nfuavtrack_2026,
210
+ title = {NF-UAVTrack: A Night-and-Fog Visible-Light Anti-UAV Tracking Dataset},
211
  author = {Cai, Zhexuan},
212
  year = {2026},
213
  howpublished = {\url{https://huggingface.co/datasets/windowflower/NF-UAVTrack}},
 
215
  }
216
  ```
217
 
218
+ The accompanying paper is under review; its reference will be added here once it is
219
+ published.
 
 
220
 
221
+ The GitHub repository <https://github.com/windowFlowers/NF-UAVTrack> hosts the
222
+ official split files and the annotation documentation.
223
 
224
  ## Contact
225
 
226
+ For questions, corrections or annotation issues, please open a discussion on this
227
+ repository.