README.md

---
license: cc-by-nc-sa-4.0
pretty_name: SPARK-2021 (SPAcecraft Recognition leveraging Knowledge of space environment)
language:
  - en
size_categories:
  - 100K<n<1M
task_categories:
  - image-classification
  - object-detection
task_ids:
  - multi-class-image-classification
tags:
  - spacecraft
  - satellite
  - space-debris
  - space-situational-awareness
  - rgb-d
  - multi-modal
  - synthetic
annotations_creators:
  - machine-generated
language_creators:
  - machine-generated
source_datasets:
  - original
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train/spark-train-*.tar
      - split: validation
        path: data/validation/spark-validation-*.tar
dataset_info:
  features:
    - name: rgb
      dtype: image
    - name: depth
      dtype: image
    - name: label
      dtype:
        class_label:
          names:
            '0': AcrimSat
            '1': Aquarius
            '2': Aura
            '3': Calipso
            '4': Cloudsat
            '5': CubeSat
            '6': Debris
            '7': Jason
            '8': Sentinel-6
            '9': Terra
            '10': TRMM
    - name: bbox
      sequence: int32
      length: 4
    - name: filename
      dtype: string
---

# SPARK-2021: SPAcecraft Recognition leveraging Knowledge of space environment

SPARK is a large-scale **multi-modal (RGB + depth) synthetic image dataset** for space object
recognition and detection, generated under a photo-realistic space simulation environment.
It was released by the [CVI² group at SnT, University of Luxembourg](https://cvi2.uni.lu/spark-2021/)
in the context of the **SPARK Challenge at IEEE ICIP 2021**.

The dataset targets **Space Situational Awareness (SSA)** applications — on-orbit servicing,
active debris removal, formation flying, and rendezvous & proximity operations — where the
scarcity of annotated spaceborne imagery is a primary bottleneck for data-driven perception.

| | |
|---|---|
| **Modalities** | RGB, depth (segmentation masks available in the original release) |
| **Images** | ~150k RGB + ~150k depth |
| **Classes** | 11 (10 satellite models + 1 combined debris class) |
| **Annotations** | Class label + 2D bounding box per image |
| **Simulator** | Unity3D, LEO scenarios around a photo-realistic Earth |
| **Type** | Fully synthetic |

---

## Dataset structure

The dataset is published as **WebDataset shards** so that it streams efficiently and pairs the
two modalities inside a single sample:

```
data/
├── train/
│   ├── spark-train-000000.tar
│   ├── spark-train-000001.tar
│   └── ...
└── validation/
    ├── spark-validation-000000.tar
    └── ...
```

Each sample inside a shard has the form:

```
<key>.rgb.jpg      # RGB image
<key>.depth.png    # 16-bit depth map, same geometry as the RGB frame
<key>.json         # {"label": 3, "class": "Calipso", "bbox": [R_min, C_min, R_max, C_max]}
```

### Splits

| Split | Samples | Notes |
|---|---|---|
| `train` | _TODO_ | Public training split of the SPARK 2021 challenge |
| `validation` | _TODO_ | Public validation split (labels released) |
| `test` | not included | Challenge test labels were kept private |

Class composition of the full release: **12,500 images per satellite class** (10 classes) and
**5,000 images per debris object** across 5 debris models, all merged into a single `Debris`
class (25,000 images) — 150,000 images in total per modality.

### Classes

| Index | Class | Type |
|---|---|---|
| 0 | AcrimSat | Satellite |
| 1 | Aquarius | Satellite |
| 2 | Aura | Satellite |
| 3 | Calipso | Satellite |
| 4 | Cloudsat | Satellite |
| 5 | CubeSat | Satellite (1RU generic CubeSat) |
| 6 | Debris | Debris (5 models merged) |
| 7 | Jason | Satellite |
| 8 | Sentinel-6 | Satellite |
| 9 | Terra | Satellite |
| 10 | TRMM | Satellite |

Satellite models come from [NASA 3D Resources](https://nasa3d.arc.nasa.gov/). Debris objects are
corrupted-texture parts of satellites and rockets: space shuttle external tank, orbital docking
system, damaged communication dish, thermal protection tiles, and connector ring.

### ⚠️ Bounding-box convention

Boxes follow the **original SPARK convention**, which is *row/column ordered*, not the usual
`x, y` ordering:

```
bbox = [R_min, C_min, R_max, C_max]   #  ==  [y_min, x_min, y_max, x_max]
```

Conversions:

```python
r_min, c_min, r_max, c_max = bbox

# Pascal VOC / torchvision  (x1, y1, x2, y2)
voc = [c_min, r_min, c_max, r_max]

# COCO  (x, y, w, h)
coco = [c_min, r_min, c_max - c_min, r_max - r_min]

# YOLO (normalised cx, cy, w, h) for an image of size (H, W)
yolo = [((c_min + c_max) / 2) / W, ((r_min + r_max) / 2) / H,
        (c_max - c_min) / W, (r_max - r_min) / H]
```

---

## Usage

```python
from datasets import load_dataset

ds = load_dataset("<org>/spark-2021", split="train")
sample = ds[0]

sample["rgb"]      # PIL.Image, RGB
sample["depth"]    # PIL.Image, 16-bit single channel
sample["label"]    # int in [0, 10]
sample["bbox"]     # [R_min, C_min, R_max, C_max]
```

### Streaming (recommended — the full dataset is large)

```python
ds = load_dataset("<org>/spark-2021", split="train", streaming=True)
for sample in ds.take(8):
    print(sample["label"], sample["bbox"], sample["rgb"].size)
```

### RGB-only classification

```python
ds = load_dataset("<org>/spark-2021", split="train").remove_columns("depth")
```

### Depth handling

Depth maps are stored as 16-bit PNGs. Convert to a float array before use:

```python
import numpy as np
depth = np.asarray(sample["depth"], dtype=np.float32)   # raw sensor units
```

Note that the released depth maps are known to be noisy and to contain holes; several challenge
entries applied morphological opening / hole filling before using them.

---

## Dataset creation

SPARK was rendered in **Unity3D**, with:

- **Earth model** — high-resolution textured 16k-polygon model based on the NASA Blue Marble
  collection, including clouds, cloud shadows, and atmospheric outer scattering.
- **Background** — high-resolution ESO panorama of the Milky Way.
- **Target** — one of the 10 satellite models or 5 debris models, randomly placed inside the
  camera field of view, in LEO.
- **Chaser** — observer platform carrying a pinhole RGB camera with known intrinsics plus a
  depth camera.

The Sun and the Earth are randomly rotated about their axes in every frame. The dataset is
deliberately spanned along four axes of variation:

1. **Scene illumination** — including extreme cases where sunlight directly faces the sensor or
   reflects off the target/Earth, producing lens flare and sensor blooming.
2. **Scene background** — Earth-in-background (rich texture, ocean/cloud specularity) vs. deep
   space (featureless, sparse stars).
3. **Range** — varying camera-to-target distance, i.e. varying target occupation of the frame.
4. **Sensor noise** — zero-mean white Gaussian noise at varying levels, emulating the high
   dynamic range and small-sensor noise of spaceborne imagers.

The baseline study in the SPARK paper found accuracy degrading systematically with lower
illumination, longer range, and increasing noise, with the **far-range + low-illumination**
subset being the hardest regime. Fine-tuning ImageNet-pretrained backbones outperformed both
random initialisation and frozen feature extraction, and RGB-D fusion reached 90.05% validation
accuracy versus 75% (RGB only) and 88.01% (depth only) at 64×64 input resolution.

---

## Original challenge protocol

The ICIP 2021 competition defined two tasks and two dedicated metrics.

**Task 1 — Classification.** Errors were weighted by severity: misclassifying a satellite as
another satellite (level 1/4), a satellite as debris (level 2/4), and — most severely — debris as
a satellite (level 4/4). Ranking used an F2-score-based metric combined with the proportion of
correctly classified non-debris samples.

**Task 2 — Detection.** Inspired by the COCO protocol: the proportion of images with both a
correct class prediction and an IoU above threshold, averaged over several IoU thresholds.

These metrics are documented here for reproducibility; this repository does not host an
evaluation server.

---

## Intended uses

- Spacecraft and debris **classification** and **detection** under space imaging conditions
- **Multi-modal RGB-D** fusion research
- **Robustness studies** with respect to illumination, range, and sensor noise
- Pretraining / representation learning for downstream proximity-operations perception

### Out of scope and limitations

- **Fully synthetic.** Models trained on SPARK alone will exhibit a substantial sim-to-real
  domain gap and should not be treated as flight-qualified without hardware-in-the-loop or
  on-orbit validation.
- **Renderer artefacts.** Illumination, flare, and noise are approximations of the true space
  radiometric environment; depth maps are simulated, not from a flight-representative sensor.
- **Class imbalance.** The single `Debris` class aggregates five geometrically distinct objects.
- **No pose labels.** SPARK provides class and bounding box only. For 6-DoF pose, see SPEED /
  SPEED+ or URSO.

---

## Citation

If you use SPARK, please cite both the dataset paper and the challenge paper:

```bibtex

@inproceedings{musallam2021sparkchallenge,
  title     = {Spacecraft Recognition Leveraging Knowledge of Space Environment:
               Simulator, Dataset, Competition Design and Analysis},
  author    = {Musallam, Mohamed Adel and Gaudilli{\`e}re, Vincent and Ghorbel, Enjie and
               Al Ismaeil, Kassem and Perez, Marcos Damian and Poucet, Michel and Aouada, Djamila},
  booktitle = {IEEE International Conference on Image Processing Challenges (ICIPC)},
  pages     = {11--15},
  year      = {2021},
  doi       = {10.1109/ICIPC53495.2021.9620184}
}
```

## Acknowledgements

Dataset produced by the Computer Vision, Imaging & Machine Intelligence (CVI²) research group,
Interdisciplinary Centre for Security, Reliability and Trust (SnT), University of Luxembourg,
in collaboration with LMO.

## Contact

Project page: <https://cvi2.uni.lu/spark-2021/>
Issues with this Hugging Face mirror: open a discussion on this repository.