File size: 9,297 Bytes
ac2467e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
175670f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ac2467e
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
---
license: apache-2.0
language:
  - zh
  - en
tags:
  - autonomous-navigation
  - drone
  - embodied-ai
  - trajectory
  - carla
  - synthetic-data
pretty_name: CosFly-Track
size_categories:
  - 100<n<1K
---

# CosFly-Track

## Overview

This is the pre-release open-source dataset for the paper **CosFly-Track: A Large-Scale Multi-Modal Dataset for UAV Visual Tracking via Multi-Constraint Trajectory Optimization**. We currently release a filtered local snapshot, and additional data will be published here in future updates.

CosFly-Track contains simulated UAV visual tracking trajectories for target tracking and waypoint prediction. Each trajectory records RGB observations, depth and segmentation artifacts, drone poses, target annotations, camera projection metadata, and navigation waypoints.

The current local snapshot contains:

- 263 parent trajectory directories.
- 526 trace directories.
- 526 `trajectory.json` files.
- Trace types observed: `ORI` and `aug_001`.
- Towns observed: `Town01`, `Town02`, `Town03`, `Town04`, `Town05`, `Town06`, `Town07`, `Town10HD` and their `_Opt` variants.

Per-trace counts:

```text
ORI      263
aug_001 263
```

Each released trajectory contains both a clean `ORI` trace and an augmented `aug_001` trace. `ORI` is the clean reference trace. `aug_001` is an augmented trace with perturbed drone pose and aligned frame indices. The augmented RGB observations and current pose can be used to train robustness, while the clean `ORI` future waypoints can be used as denoising supervision when desired.

## Directory Layout

The filtered upload data is organized by town, parent trajectory, and trace type. The original date level has been removed from the release layout.

```text
CosFly-Track/
├── README_ZH.md
├── data_sample/
│   └── trajectory_<id>/
│       ├── ORI/
│       │   ├── trajectory.json
│       │   └── frames_playback/
│       │       └── frame_<index>/
│       │           ├── rgb.png
│       │           ├── depth.npy
│       │           ├── instance.png
│       │           ├── debug.png
│       │           └── meta.json
│       └── aug_001/
│           ├── trajectory.json
│           ├── perturbation_report.json
│           └── frames_playback/
│               └── frame_<index>/
│                   ├── rgb.png
│                   ├── depth.npy
│                   ├── instance.png
│                   ├── debug.png
│                   └── meta.json
└── data_v7/
    └── Town<id>/
        └── trajectory_<id>/
            ├── augmentation_summary.json
            ├── ORI/
            │   ├── trajectory.json
            │   └── frames_playback/
            │       └── frame_<index>/
            │           ├── rgb.png
            │           ├── depth.npy
            │           ├── instance.png
            │           ├── debug.png
            │           └── meta.json
            └── aug_001/
                ├── trajectory.json
                ├── perturbation_report.json
                └── frames_playback/
                    └── frame_<index>/
                        ├── rgb.png
                        ├── depth.npy
                        ├── instance.png
                        ├── debug.png
                        └── meta.json
```

## Data Files

Each trace contains one trajectory-level JSON file:

```text
data_v7/<Town>/trajectory_<id>/<trace_dir>/trajectory.json
```

Example:

```text
data_v7/Town01/trajectory_1777083733/aug_001/trajectory.json
```

Observed top-level keys include:

```text
camera
dataset_format
frames_dir
pedestrian_blueprint
points
schema
schema_version
source
trace_dir
```

Additional augmentation metadata is stored at the parent trajectory level:

```text
data_v7/<Town>/trajectory_<id>/augmentation_summary.json
data_v7/<Town>/trajectory_<id>/aug_001/perturbation_report.json
```

The `points` array stores per-frame annotations. Observed per-point keys include:

```text
drone_pose
index
is_perturbed
nav_waypoint
perturbation
target
timing
world_to_camera
```

Important nested fields:

- `drone_pose`: UAV position and attitude, including `x`, `y`, `z`, `pitch`, `yaw`, and `roll`.
- `target`: target actor metadata and visual annotations, including visibility, image coordinates, depth, and 3D bounding box.
- `nav_waypoint`: navigation waypoint annotations in both world and image coordinates.
- `world_to_camera`: transformation matrix for projecting world coordinates to the camera frame.

Each playback frame directory contains:

- `rgb.png`: RGB observation, usually 1280 x 720.
- `depth.npy`: depth array in NumPy format.
- `instance.png`: instance segmentation image.
- `debug.png`: visualization/debug image with overlays.
- `meta.json`: per-frame metadata aligned with the corresponding `points` entry.

## Minimal Loader Example

```python
from pathlib import Path
import json
import numpy as np
from PIL import Image

root = Path("data_v7")

for traj_json in sorted(root.glob("Town*/trajectory_*/*/trajectory.json")):
    with traj_json.open("r", encoding="utf-8") as f:
        traj = json.load(f)

    trace_dir = traj.get("trace_dir")
    points = traj.get("points", [])
    if not points:
        continue

    first = points[0]
    frame_index = first["index"]
    frame_dir = traj_json.parent / "frames_playback" / f"frame_{frame_index:05d}"

    rgb = Image.open(frame_dir / "rgb.png")
    depth = np.load(frame_dir / "depth.npy")
    with (frame_dir / "meta.json").open("r", encoding="utf-8") as f:
        meta = json.load(f)

    print(
        trace_dir,
        len(points),
        rgb.size,
        depth.shape,
        first["drone_pose"],
        first["target"],
        meta["frame_id"],
    )
    break
```

## Filtering and Quality Criteria

The current release is generated from filtered v7 data. Main checks include:

- Integrity checks: `ORI` and `aug_*` traces exist, required frame files are readable, and `meta.json` contains `frame_id`.
- Basic trajectory quality: average target-drone distance <= 35, max distance <= 50, max target height <= 2, target visible ratio > 55%, and adjacent drone z-step <= 5.
- Drone collision checks: drone poses must not enter town-specific map 3D bounding boxes.
- Target collision checks: target 3D bounding boxes must not overlap map objects beyond the configured threshold.
- Town-map consistency: map bounding boxes are selected according to the Town name in each trajectory path.

## Recommended Quality Checks

Before packaging a public release, run the following checks:

- Ensure every parent trajectory contains both `ORI` and `aug_001`.
- Ensure every trace has a readable `trajectory.json`.
- Ensure every frame referenced by `trajectory.json` has `meta.json`, `rgb.png`, `depth.npy`, `instance.png`, and `debug.png`.
- Optionally decode all RGB and debug images with Pillow to catch corrupted images.
- Verify that frame indices between `ORI` and `aug_001` are aligned.
- Verify that public manifests and JSON metadata do not expose unintended internal machine paths.

## Coordinate System

The trajectory fields `x`, `y`, `z`, `pitch`, `yaw`, and `roll` follow CARLA / Unreal-style world coordinates and Euler angles. For Three.js visualization, the project uses the mapping `(x, z, -y)`.

## Intended Uses

CosFly-Track can be used for:

- UAV visual target tracking.
- Waypoint prediction and visual navigation.
- Robustness training with augmented UAV poses.
- Multi-modal learning from RGB, depth, segmentation, and structured trajectory metadata.
- Trajectory prediction, visibility modeling, and simulator-based evaluation.

## Release Checklist

The following items should be finalized before public open-source release:

- Public Hugging Face dataset URL.
- Dataset license.
- Paper authors and final citation.
- Train/validation/test split manifest.
- Benchmark metric definition and evaluation script path.
- Known limitations and simulator domain-gap notes.

## Citation

Citation information is pending. Replace this section with the final BibTeX entry before release.

```bibtex
@misc{chen2026cosflyplanmatrixfly,
      title={CosFly: Plan in the Matrix, Fly in the World}, 
      author={Hanxuan Chen and Xiangyue Wang and Songsheng Cheng and Ruilong Ren and Jie Zheng and Shuai Yuan and Tianle Zeng and Hanzhong Guo and Binbo Li and Kangli Wang and Ji Pei},
      year={2026},
      eprint={2605.19120},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2605.19120}, 
}
@misc{wang2026cosflytracklargescalemultimodaldataset,
      title={CosFly-Track: A Large-Scale Multi-Modal Dataset for UAV Visual Tracking via Multi-Constraint Trajectory Optimization}, 
      author={Xiangyue Wang and Hanxuan Chen and Songsheng Cheng and Ruilong Ren and Jie Zheng and Shuai Yuan and Tianle Zeng and Hanzhong Guo and Kangli Wang and Ji Pei},
      year={2026},
      eprint={2605.17776},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2605.17776}, 
}
```