README.md

---
license: other
license_name: reap-proprietary-v1
license_link: LICENSE
pipeline_tag: object-detection
tags:
  - object-detection
  - instance-segmentation
  - thermal
  - infrared
  - counter-uas
  - c-uas
  - sim-to-real
  - defense
  - military
---

# REAP-T2 "Scythe"

**Thermal counter-UAS detector and instance segmenter.** REAP: Recognition and
Engagement of Aerial Platforms.

REAP-T2 detects and type-classifies fixed-wing and loitering-munition UAVs in
single-frame thermal imagery. The model was trained on synthetic data
exclusively. It has never seen a real thermal frame. On real footage of
Shahed-136 targets it holds an 88% center-hit rate at 0.28 false positives per
frame. It ships with `trakk`, a real-time tracker written in Rust, and the two
run as one system.

Weights are not distributed. REAP-T2 is a military product under Romanian
export control. This card documents the system. Partnerships:
**office@187.ro**.

<p align="center">
  <img src="https://blog.187.ro/assets/reap_demo.gif" alt="REAP-T2 + trakk running on real footage, four 3s excerpts" width="360">
</p>

*REAP-T2 → `trakk` on real footage: boxes, gate score, and track ID as rendered
by the pipeline. Excerpts used for analysis and demonstration, ~3 s each,
credited below. Original footage © the respective owners.*

**Sources:**
1. [Wild Hornets](https://wildhornets.com/en/)
2. [ZRK](https://www.zrk.tech/)

Note: We are not affiliated with the above companies. 

---

## Model details

- **Developed by:** UNUOPTSAPTE SRL (DBA 187DEF)
- **Version:** REAP-T2 ("Scythe")
- **Type:** DETR-family detector with an instance segmentation head
- **Input:** single-frame thermal imagery
- **Outputs:** bounding boxes, instance masks, a per-detection binary presence
  ("gate") score, and a fine-class label
- **Classes:** 7 fixed-wing and loitering-munition UAV types, listed below
- **Training data:** 100% synthetic thermal imagery. Zero real frames.
- **License:** REAP Proprietary License v1 (`reap-proprietary-v1`), export
  controlled
- **Bundled tracker:** `trakk`, real-time track-by-detection, Rust

---

## Detected classes

| Class | Category |
|---|---|
| Shahed-136 | One-way-attack loitering munition |
| Gerbera | Decoy / multirole UAV |
| Lancet-3 | Loitering munition |
| Granat-4 | Reconnaissance UAV |
| Supercam S350 | Reconnaissance UAV |
| Orlan-10 | Reconnaissance UAV |
| ZALA 421-16 | Reconnaissance UAV |

---

## Intended use

Per-frame thermal detection of UAV threats for counter-UAS situational
awareness and engagement support.

The detector feeds the bundled tracker, and the two should be read as one
system. `trakk` holds a target through frames the detector misses, decides a
track's class by voting across its whole lifetime, and discards detections that
never persist into a stable track. Track-level performance runs well above the
per-frame numbers reported below. The gate score is the engagement-lock signal.

### Prohibited

Any use outside the terms of the REAP Proprietary License v1. Outputs are
decision support for a human operator and must never be wired to an autonomous
engagement authority.

---

## Military product & export control

REAP-T2 is a military counter-UAS product built to counter one-way-attack UAV
and loitering-munition threats. It is subject to Romanian export control
legislation, administered by ANCEX. Any transfer of the model, weights, or
associated technology requires authorization under that framework. UNUOPTSAPTE
SRL (DBA 187DEF) does not authorize offensive use, use against non-combatants,
or use as an unsupervised autonomous weapon. Human-in-the-loop engagement
decisions are assumed.

---

## Evaluation

**Bench-T1:** ~3000 real, hard thermal frames of Shahed-136 targets, captured
independently of any training data. No real footage of any kind was used in
training, so every number below is a pure sim-to-real transfer result.

Standard IoU-averaged COCO metrics understate detection performance here
because the model and the real-footage annotations follow different
bounding-box conventions. We report metrics that measure whether the target was
found, typed, and how often the model cries wolf:

- **Center-hit (IoU ≥ 0.3):** a predicted box landed on the target,
  independent of box size.
- **Corrected recall @ 0.5:** recall at IoU 0.5 with predicted boxes
  normalized to the annotation convention.
- **Typing:** fraction of center-hits assigned the correct fine class.
- **FP/frame:** false positives per frame at the operating point.

### Results (per-frame, single checkpoint)

| Metric | Value |
|---|---|
| Center-hit (IoU ≥ 0.3) | 88% |
| Corrected recall @ 0.5 | 81% |
| Fine-typing accuracy | 53% |
| FP / frame | 0.28 |

These are single-frame numbers from a model that has never seen reality. In
deployment, per-frame typing is the floor: `trakk` votes class across a
track's entire lifetime, so no single frame's label is load-bearing, and the
gate signal carries the engagement decision from first lock.

---

## Bundled tracker: `trakk`

`trakk` is a real-time track-by-detection pipeline that consumes the detector's
boxes, gate scores, and class logits. It provides track-level recall recovery,
temporal class fusion, false-positive suppression via track persistence, and an
engagement-lock signal.

### Real-footage pipeline evaluation

The full REAP-T2 → `trakk` pipeline, run on real operational footage. These are
behavioral metrics with no ground-truth track labels: they characterize
end-to-end behavior in the wild and complement the labeled Bench-T1 detector
metrics above.

| Metric | Thermal / night (seeker) | Daylight EO (interceptor) |
|---|---|---|
| Footage | 2929 fr / 98 s @ 30 fps | 2091 fr / 70 s @ 30 fps |
| Detector high-conf (≥0.5) hit rate | 54% of frames | 45% of frames |
| **Track coverage** (confirmed track present) | **90% of frames** | **100% of frames** |
| First-lock latency | **530 ms** | **370 ms** |
| Engagement-lock uptime | 89% | 97% |
| Longest continuous track | 43.1 s | 24.6 s |

Track coverage runs at roughly double the per-frame detector hit rate on both
clips: `trakk` stitches sparse detections into near-continuous tracks. First
lock arrives in well under a second, and engagement-lock holds for the large
majority of each engagement.

The daylight-EO clip is worth reading twice. REAP-T2 was trained on synthetic
thermal imagery, so daylight EO is two domain jumps away from anything it was
trained on, and the pipeline still confirmed a track in 100% of frames and held
engagement-lock for 97% of a 70-second engagement.

---

## Known behavior & envelope

Reported detector metrics are per-frame; the deployed system is detector plus
tracker, and the pipeline evaluation above shows how far end-to-end behavior
runs ahead of single-frame numbers. Performance is characterized within the
range and viewpoint envelope represented in training and evaluation, and
validation across a broader envelope, longer video, and cluttered scenes with
non-drone distractors is in progress. Training data remains 100% synthetic by
design: every improvement to sensor and signature fidelity in the data pipeline
lands directly on real-footage numbers, and that work is ongoing.

---

## Access

Partnerships and evaluations: **office@187.ro**

All work, including the REAP-T2 model, weights, checkpoints, training data, the
`trakk` tracker, and this documentation, is © 2026 **UNUOPTSAPTE SRL
(DBA 187DEF)**, all rights reserved. Portions of the detector build on RF-DETR
(© Roboflow, Inc.), which remains subject to its upstream **Apache License
2.0**; the REAP Proprietary License v1 governs UNUOPTSAPTE SRL's modifications,
weights, datasets, and the bundled tracker.

---

## Citation

```bibtex
@misc{reap-t2-2026,
  title  = {REAP-T2 "Scythe": A Synthetic-Trained Thermal Counter-UAS Detector},
  author = {{UNUOPTSAPTE SRL (DBA 187DEF)}},
  year   = {2026},
  note   = {Military product. Weights not distributed. Export controlled (Romania / ANCEX).}
}
```

---

*Model version REAP-T2 "Scythe" · Card revision 2026-07-14.*