README.md

---
license: apache-2.0
task_categories:
  - image-segmentation
  - object-detection
tags:
  - image-segmentation
  - instance-segmentation
  - drone
  - uav
  - robotics
  - sae-eletroquad
size_categories:
  - 10K<n<100K
pretty_name: SAE Eletroquad 2026 - Hook (Hang the Wire)
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
dataset_info:
  features:
    - name: image
      dtype: image
    - name: image_id
      dtype: int64
    - name: width
      dtype: int32
    - name: height
      dtype: int32
    - name: objects
      struct:
        - name: id
          sequence: int64
        - name: bbox
          sequence:
            sequence: float32
            length: 4
        - name: category
          sequence:
            class_label:
              names:
                '0': rose
                '1': sphere
        - name: area
          sequence: float64
        - name: segmentation
          sequence:
            sequence: float32
---

# SAE Eletroquad 2026 - Hook (Hang the Wire)

Instance-segmentation dataset for the "hang the wire" hook mission of the SAE
Brasil Eletroquad 2026 competition, collected and annotated by
[Black Bee Drones](https://github.com/Black-Bee-Drones). Two classes: `rose`
(each visible segment of the suspended rope) and `sphere` (the orange sphere
mounted on the rope).

**Trained model:** [blackbeedrones/sae-2026-hang-all-yolo26n-seg-v2-960](https://huggingface.co/blackbeedrones/sae-2026-hang-all-yolo26n-seg-v2-960) — a YOLO26n-seg model trained and evaluated on this exact version of the data.

## Competition

[SAE Brasil Eletroquad](https://saebrasil.org.br/programas-estudantis/eletroquad/)
is a student competition for autonomous quadrotors. The 2026 edition ran on
14–17 May 2026 at Univap (São José dos Campos, SP). Black Bee Drones finished
2nd overall ([official results](https://arquivos.saebrasil.org.br/2026/EletroQuad/EletroQuad2026ResultadoOficial.pdf)).

## Mission

In the hook mission the drone takes off, finds the orange sphere mounted on one
of two suspended ropes, parks a fixed real-world distance from it, picks which
side of the rope to fly along, turns perpendicular to the rope, descends on
LIDAR, releases a hook with a servo, and lands. A single YOLO-seg model runs on
every control-loop tick: each visible rope segment is its own `rose` instance,
which is what lets the controller choose a side. The mission code is the
[`hook`](https://github.com/Black-Bee-Drones) ROS 2 package, built on
[Nectar SDK](https://github.com/Black-Bee-Drones/nectar-sdk).

## Classes

| id | name | description |
|----|------|-------------|
| 0 | `rose` | A single visible segment of the suspended rope/wire. Several instances can appear per image. |
| 1 | `sphere` | The orange sphere on the rope; the mission's primary visual anchor. |

## Dataset structure

| Split | Images | `rose` | `sphere` | Instances |
|-------|-------:|-------:|---------:|----------:|
| train | 9,235 | 7,787 | 3,402 | 11,189 |
| validation | 396 | 355 | 154 | 509 |
| test | 395 | 354 | 147 | 501 |
| **Total** | **10,026** | **8,496** | **3,703** | **12,199** |

Each row is an image plus an `objects` struct with one entry per instance:

- `bbox` — axis-aligned box `[x_min, y_min, width, height]` in absolute pixels, computed from the polygon.
- `area` — polygon area in pixels.
- `category` — `ClassLabel` (`rose` / `sphere`).
- `segmentation` — a flat polygon `[x1, y1, x2, y2, ...]` in absolute pixels (one polygon per instance).

Image dimensions are stored per row in `width` / `height`. The Hub viewer draws
the bounding boxes; the polygons are kept for mask training.

## Provenance

Exported from the Roboflow project `black-bee-drones/sae-2026-hang` (version 2)
and mirrored here. This is the version the published model was trained and
evaluated on — its test split (395 images, 501 instances) matches the model's
reported evaluation.

## Usage

### Load with HuggingFace Datasets

```python
from datasets import load_dataset

dataset = load_dataset("blackbeedrones/sae-2026-hook")
example = dataset["train"][0]
print(example["objects"]["category"], example["objects"]["segmentation"][0][:8])
```

### Materialize as YOLO-seg (for training)

```python
from nectar.ai.segmentation.datasets import HuggingFaceSegHandler

handler = HuggingFaceSegHandler("data/sae-2026-hook")
handler.download(repo_id="blackbeedrones/sae-2026-hook", format_type="yolo")
# data/sae-2026-hook now has data.yaml + {train,valid,test}/{images,labels}
```

Or from the command line:

```bash
nectar-ai segment dataset download --source huggingface \
  --repo blackbeedrones/sae-2026-hook --format yolo --output data/sae-2026-hook
```

### Train with Nectar SDK

```python
from nectar.ai.segmentation import Segmentor, SegTrainingConfig

segmentor = Segmentor("yolo26n-seg.pt")
segmentor.load()
segmentor.train(SegTrainingConfig(dataset_path="data/sae-2026-hook/data.yaml", epochs=100, imgsz=960))
```

The published model runs at `imgsz=960`, `iou=0.6`, with per-class confidence
thresholds `rose=0.47` and `sphere=0.70`.

## References

- Trained model: [blackbeedrones/sae-2026-hang-all-yolo26n-seg-v2-960](https://huggingface.co/blackbeedrones/sae-2026-hang-all-yolo26n-seg-v2-960)
- [Nectar SDK](https://github.com/Black-Bee-Drones/nectar-sdk)
- [SAE Brasil Eletroquad](https://saebrasil.org.br/programas-estudantis/eletroquad/)
- [Eletroquad 2026 official results](https://arquivos.saebrasil.org.br/2026/EletroQuad/EletroQuad2026ResultadoOficial.pdf)