CVI2-UniLU/SPARK-2024

SPARK 2024 — Stream 2: Spacecraft Trajectory Estimation

This dataset comes from Stream 2 of the SPARK 2024 challenge (SPAcecraft Recognition leveraging Knowledge of space environment). The task is 6-DoF pose / trajectory estimation of a target spacecraft from monocular camera images: given a sequence of images, estimate the relative position (translation) and orientation (quaternion) of the spacecraft with respect to the camera at every frame.

Dataset Summary

Split	Sequences	Images	Size
Train	100 (RT500–RT599)	30,000	~6.8 GB
Test	4 (RT001–RT004)	2,123	~600 MB

• Image format: JPEG, 1440 × 1080, RGB
• Labels: per-frame 6-DoF pose — translation (Tx, Ty, Tz) in meters and orientation as a unit quaternion (Qx, Qy, Qz, Qw) (scalar-last convention, compatible with scipy.spatial.transform.Rotation)
• Camera intrinsics: provided in K.txt (3×3 matrix)
• Mockup scale: 2.5

Dataset Structure

.
├── K.txt                  # Camera intrinsic matrix (3×3)
├── train.csv              # Pose labels for all training images
├── test.csv               # Pose labels for all test images
├── train/                 # 100 trajectory folders
│   ├── RT500/
│   │   ├── img000_RT500.jpg
│   │   ├── img001_RT500.jpg
│   │   └── ...
│   └── ...
├── test/                  # 4 trajectory folders
│   ├── RT001/
│   └── ...
├── visualize_data.py      # Script to project pose axes onto images
└── requirements.txt       # Dependencies for the visualization script

Annotations (train.csv / test.csv)

Each row describes one image:

Column	Description
filename	Image file name, e.g. img000_RT590.jpg
sequence	Trajectory ID, e.g. RT590 (also the folder name)
Tx, Ty, Tz	Translation of the spacecraft in the camera frame (meters)
Qx, Qy, Qz, Qw	Orientation quaternion (scalar-last)

Camera Intrinsics (K.txt)

[[1744.92206139719, 0, 720.0],
 [0, 1746.58640701753, 540.0],
 [0, 0, 1]]

Usage

Download

pip install huggingface_hub
hf download <username>/spark2024-stream-2 --repo-type dataset --local-dir spark2024-stream-2

Load the labels and an image

import numpy as np
import pandas as pd
from PIL import Image

df = pd.read_csv("train.csv")
row = df.iloc[0]

img = Image.open(f"train/{row.sequence}/{row.filename}")
translation = row[["Tx", "Ty", "Tz"]].to_numpy(dtype=float)
quaternion = row[["Qx", "Qy", "Qz", "Qw"]].to_numpy(dtype=float)  # scalar-last

with open("K.txt") as f:
    K = np.array(eval(f.read()))

Visualize ground-truth poses

visualize_data.py projects the spacecraft body axes onto random training images using the ground-truth pose and the camera intrinsics:

pip install -r requirements.txt
python visualize_data.py

Intended Uses

• 6-DoF spacecraft pose estimation from monocular images
• Sequential / trajectory-based pose estimation and filtering
• Domain adaptation and sim-to-real research for space imagery

Citation

If you use this dataset, please cite the SPARK challenge organized by the CVI² group at SnT, University of Luxembourg:

@dataset{rathinam_2024,
  author       = {Rathinam, Arunkumar and
                  Mohamed Ali, Mohamed Adel and
                  Gaudilliere, Vincent and
                  Aouada, Djamila},
  title        = {SPARK 2024: Datasets for Spacecraft Semantic
                   Segmentation and Spacecraft Trajectory Estimation
                  },
  month        = feb,
  year         = 2024,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.10908215},
  url          = {https://doi.org/10.5281/zenodo.10908215},
}