waltgrace/fiber-optic-drones

---
license: cc-by-nc-4.0
task_categories:
- object-detection
- image-classification
language:
- en
tags:
- drones
- fiber-optic
- object-detection
- vlm-labeled
- data-label-factory
size_categories:
- 1K<n<10K
pretty_name: Fiber-Optic Drones (full)
configs:
- config_name: default
data_files:
- split: train
path: data.parquet
---

# Fiber-Optic Drones (full release)

**2,260 images + 8,759 bounding boxes + 5,114 verified detections** for fiber-optic drone object detection. Image bytes are inlined so you can `load_dataset()` and start training immediately.

If you only want labels (no pixels, fully redistributable), see the sister dataset:
[`waltgrace/fiber-optic-drones-labels`](https://huggingface.co/datasets/waltgrace/fiber-optic-drones-labels).

## What's in here

- **2,260** images bundled inline (~309 MB Parquet)
- **8,759** bounding boxes (Falcon Perception)
- **5,114** boxes (58%) verified YES by Qwen 2.5-VL-3B
- **5 categories**: `fiber optic spool`, `cable spool`, `drone`, `quadcopter`, `fiber optic drone`
- **5 buckets** (gather sources): `positive/fiber_spool_drone`, `positive/spool_only`, `negative/drones_no_spool`, `distractor/round_things`, `background/empty`

## Quick start

```python
from datasets import load_dataset

ds = load_dataset("waltgrace/fiber-optic-drones", split="train")
print(ds)
# Dataset({
# features: ['image', 'image_id', 'file_name', 'bucket', 'width', 'height',
# 'n_bboxes', 'n_approved', 'bboxes'],
# num_rows: 2260
# })

row = ds[0]
img = row["image"] # PIL.Image.Image
print(img.size) # (640, 360)
print(row["bucket"]) # "positive/fiber_spool_drone"
print(row["n_bboxes"], "boxes") # 12 boxes

# Bboxes are stored as struct-of-lists for fast columnar access:
for ann_id, cat, x1, y1, x2, y2, verdict in zip(
row["bboxes"]["annotation_id"],
row["bboxes"]["category"],
row["bboxes"]["x1"], row["bboxes"]["y1"],
row["bboxes"]["x2"], row["bboxes"]["y2"],
row["bboxes"]["vlm_verdict"],
):
print(f" ann {ann_id}: {cat} ({x1:.0f},{y1:.0f},{x2:.0f},{y2:.0f}) → {verdict}")
```

Bbox coordinates are **pixel space** (not normalized), origin top-left.

## Filtering examples

```python
# Only positive-bucket images that have at least one approved bbox
positives = ds.filter(
lambda r: r["bucket"].startswith("positive/") and r["n_approved"] > 0
)

# Only YES-verified boxes for a specific category
def keep_yes_drones(row):
new_b = {k: [] for k in row["bboxes"]}
for i in range(row["n_bboxes"]):
if (row["bboxes"]["category"][i] == "drone"
and row["bboxes"]["vlm_verdict"][i] == "YES"):
for k in new_b:
new_b[k].append(row["bboxes"][k][i])
return {**row, "bboxes": new_b, "n_bboxes": len(new_b["annotation_id"])}

drones_only = ds.map(keep_yes_drones)
```

## How was this labeled?

Two-stage local pipeline running on a 16 GB Apple Silicon Mac:

1. **Falcon Perception** (TII, ~600 MB) drew 8,759 candidate bounding boxes across 2,260 web-scraped images using 5 query prompts.
2. **Qwen 2.5-VL-3B-Instruct** (Alibaba, ~2.5 GB) cropped each bbox + context and answered "Is this a `<category>`? YES / NO / UNSURE" with brief reasoning.

The full pipeline (gather → filter → label → verify → review) is open source:
[`walter-grace/data-label-factory`](https://github.com/walter-grace/data-label-factory).

Reproduce in five commands:

```bash
pip install git+https://github.com/walter-grace/data-label-factory
python3 -m mlx_vlm.server --model mlx-community/Qwen2.5-VL-3B-Instruct-4bit --port 8291
# (start mac_tensor with --vision --falcon for the label stage)
data_label_factory pipeline --project projects/drones.yaml
```

## Per-query agreement (Falcon ↔ Qwen)

| Query | Falcon detections | Qwen YES rate |
|---|---:|---:|
| cable spool | 2,798 | 88% |
| quadcopter | 1,805 | 81% |
| drone | 2,186 | 80% |
| fiber optic drone | 573 | 77% |
| fiber optic spool | 1,397 | 57% |

`fiber optic spool` is the niche query — Falcon overfires, Qwen rejects 43%.

## License (READ THIS)

**License: CC-BY-NC 4.0 — research / non-commercial use only.**

The 2,260 images were gathered from DuckDuckGo, Wikimedia Commons, Openverse,
and YouTube. Original copyright belongs to the individual creators of each
source. They are bundled here for **research purposes** under fair-use carve-outs
that allow educational and non-commercial ML research.

If you intend to use this dataset for **commercial** purposes (training a
production model, building a SaaS, etc.), you must:
- Contact the original image creators for permission, OR
- Re-gather images yourself using the labels-only release at
[`waltgrace/fiber-optic-drones-labels`](https://huggingface.co/datasets/waltgrace/fiber-optic-drones-labels)
combined with your own image sourcing.

The **labels themselves** (bboxes, categories, VLM verdicts, reasoning) are
released under Apache 2.0 and are unrestricted.

If you are a copyright holder of any image in this dataset and want it removed,
open an issue on the repo or email the maintainer — we will remove it immediately.

## Citation

```bibtex
@dataset{walter-grace-2026-fiber-optic-drones,
author = {walter-grace},
title = {Fiber-Optic Drones},
year = 2026,
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/waltgrace/fiber-optic-drones},
}
```