---
license: cc-by-4.0
task_categories:
  - object-detection
  - keypoint-detection
  - robotics
tags:
  - spacecraft
  - spacecraft-detection
  - spacecraft-pose-estimation
  - pose-estimation
  - event-sensing
  - event-camera
  - neuromorphic
  - computer-vision
pretty_name: SPADES
size_categories:
  - 100K<n<1M
viewer: false
---

## SPADES dataset

**SPADES - SPAcecraft Pose Estimation Dataset using Event Sensing**, a unique and new space dataset designed to advance spacecraft pose estimation research. SPADES dataset contains two categories of data: **Synthetic** and **Real**. 

**Synthetic dataset** focuses on simulating RGB images of a satellite target—in this case, **Proba-2**—by moving a spacecraft model along predefined trajectories within the simulator’s camera field of view. To generate realistic imagery, the **Unreal Engine** (UE) was used as a simulation environment, various camera orientations and distances were simulated. Additionally, the background included dynamic representations of the Sun and Earth, both of which were animated to rotate around their respective axes according to each trajectory.  The dataset also includes event data generated by the ICNS event simulator, which models the behavior of neuromorphic sensors. Blender was used to simulate sensor behavior, producing an event stream from the RGB images generated in the UE environment. 

The **synthetic dataset** consists of 300 trajectories, each with approximately 600 RGB images per trajectory, along with corresponding pose labels. In total, the dataset includes 179,400 pose labels. Event streams for each trajectory are stored in separate CSV files, with timestamps matching the pose labels, ensuring a precise time sequence.

**Real data** was collected at the Zero-G Laboratory at the SnT, University of Luxembourg. A scaled mockup of the Proba-2 satellite was used for data collection, with event data captured by a **Prophesee Metavision EVK4-HD** equipped with the SONY IMX636ES (HD) event vision sensor. 

# Dataset Description

This dataset contains both **synthetic** and **real** event streams, along with their corresponding ground-truth labels. The data is organized in directories for events and labels for both types of datasets.

## Dataset Structure - Synthetic

```
Synthetic
│
├── RT000.h5
├── RT001.h5
├── RT002.h5
└── ...
```

**Data Format (HDF5)**: Each `.h5` file represents a stream of synthetic events with relevant ground-truth pose information. Each sequence (e.g., `RT000.h5`) containing two main groups:

```
/events
/labels
```

Relation Between Events and Labels

- Event data provides **asynchronous, high-frequency motion cues**
- Ground-truth labels provide **discrete pose** at fixed intervals
- Events occurring within:`[timestamp_i, timestamp_{i+1}]` correspond to the spacecraft motion between two consecutive poses.

#### 1. Event Stream Data (`/events` group)

The **event stream** is stored in the `/events` group :

| Dataset | Type | Description |
|--------|------|-------------|
| `xs` | int | X-coordinate of the event on the image plane |
| `ys` | int | Y-coordinate of the event on the image plane |
| `ts` | int | Timestamp of the event |
| `ps` | int | Event polarity (ON / OFF) |

**Event Semantics**

- **Timestamp unit**:    `ts` is expressed in **100 microseconds (µs)**.
- **Time range**:    Starts at `0` and ends at approximately `600000`, corresponding to:
600000 × 100 µs = 60,000,000 µs = 60 s
- **Spatial coordinates**:  `(xs, ys)` represent the 2D pixel location of each event.
- **Polarity**: `ps = 1` → **ON** event  and  `ps = 0` → **OFF** event



#### 2. Ground-Truth Labels (`/labels/data` dataset)

Ground-truth pose annotations are stored as a **structured dataset** at:

```
/labels/data
```
Each entry corresponds to one spacecraft pose.

**Label Fields**

| Field | Type | Description |
|------|------|-------------|
| `filename` | string | RGB image filename (e.g., `img000_RT000.png`) |
| `Tx, Ty, Tz` | float | 3D translation vector (spacecraft position) |
| `Qx, Qy, Qz, Qw` | float | Quaternion representing spacecraft orientation |
| `timestamp` | int | Timestamp of the pose |

#### Timestamp Semantics

- The first label (`img000_RT000.png`) corresponds to **0 s**
- Each subsequent label is offset by **100 ms**
- The `timestamp` field uses the **same 100 µs timebase** as `/events/ts`


#### 3. Pose Timing and Sampling

- **Number of poses**: `599`
- **Sampling rate**: `10 Hz`
- **Time interval**: `0.1 s` (100 ms)
- **Total duration**: `~60 s`



---

### Real Dataset (Test data)


```
Real
│
├── RT901.h5
├── RT902.h5
├── RT903.h5
└── ...
```

**Data Format (HDF5)**: Each `.h5` file represents a stream of real events. Each sequence (e.g., `T000.h5`) contains only `/events`.

- Event Stream Data (`/events` group) is structured similar to synthetic data with the following difference in timestamps.

- **timestamp**: The timestamp of the event, in **1 µs**. For real data, the range starts at `0` and ends at approximately `50,000,000 μs` (representing a duration of **50 s**).

---

### Citation

If you use this dataset in your research or project, please cite it as follows:

```
@INPROCEEDINGS{10611231,
author={Rathinam, Arunkumar and Qadadri, Haytam and Aouada, Djamila},
booktitle={2024 IEEE International Conference on Robotics and Automation (ICRA)}, 
title={SPADES: A Realistic Spacecraft Pose Estimation Dataset using Event Sensing}, 
year={2024},
pages={11760-11766},
doi={10.1109/ICRA57147.2024.10611231}}
```

---