18,477source frames7 bags · 30 fps
1,235frames sampled6.68% of corpus · 2 fps
96ground truth frames72 drone · 24 drone-free
10m 15.6srecordedone 24.4-min session
20.6 GBuncompressed mcap0.98–1.24 MB / frame
30–22,673 px²drone size0.009–5.5% of frame
Where these frames come from
Seven ROS 2 bags of white-hot thermal video from a single counter-UAV session — four held
out for evaluation, three used to fine-tune the detector.
- Container
- rosbag2 v5 · mcap, no compression
- Stream
- One topic — /image_raw
(sensor_msgs/msg/Image, CDR, frame_id = camera_frame).
No camera_info, no IMU, no telemetry.
- Encoding
- bgr8 — 3 × 8-bit, step = width×3.
A colour container holding greyscale: attack1 is bit-exact R = G = B,
the rest carry ≤3 levels of chroma residue from a YUV round-trip. Platform colourfulness 1.1; mean
R/G/B 0.319 / 0.327 / 0.316.
- Signal
- LWIR, white-hot — targets read brighter than sky (GT-box peak 189–255 against frame
means of 78–120).
- Cameras
- Fixed 640×512 cores, three of them stored rotated to portrait 512×640; handheld rigs at
720×576 (PAL D1) with an OSD band burnt into the pixels.
- Sampling
- Every 15th frame reviewed → 1,235 candidates → 96 pinned here.
| Recording | Role | Geometry | Frames | Duration | mcap | Drone samples | Drone-free samples |
|---|
| attack1 | held out | 640×512 | 3,484 | 1m 56.1s | 3.43 GB | 18 | 6 |
| cam3_1 | held out | 512×640 | 1,577 | 52.5s | 1.55 GB | 18 | 6 |
| uav_20250405_1 | held out | 720×576 | 2,069 | 1m 08.9s | 2.57 GB | 18 | 6 |
| uav_20250405_2 | held out | 720×576 | 2,542 | 1m 24.7s | 3.16 GB | 18 | 6 |
| cam1_1 | fine-tune | 512×640 | 2,563 | 1m 25.4s | 2.52 GB | 68 | 22 |
| cam2_1 | fine-tune | 512×640 | 1,609 | 53.6s | 1.58 GB | 43 | 14 |
| uav_20250405_4 | fine-tune | 720×576 | 4,633 | 2m 34.4s | 5.77 GB | 104 | 34 |
| 7 bags | | | 18,477 | 10m 15.6s | 20.58 GB | 287 | 94 |
No radiometry survives. These are already palette-rendered 8-bit frames, not raw 16-bit
sensor counts, and nothing in the bags carries gain or offset — temperature is unrecoverable, so a drone is
only ever a bright blob against sky. That is why a 4 px speck is a genuinely hard target.
One session, not one dataset. All seven bags were recorded inside a single 24.4-minute
window on 12 Aug 2026 (03:32–03:57). The four held-out recordings share site, lighting and airframe
with the three used for fine-tuning, so scores here characterise in-domain performance — not
generalisation to a new environment.
