zhixiang320/Spacecraft_Detection_Keypoint_and_Pose_Dataset

Dataset Card: SDKP (Spacecraft Detection, Keypoint, and Pose Dataset)

Dataset Summary

The SDKP (Spacecraft Detection, Keypoint, and Pose) dataset is a comprehensive benchmark designed to advance spacecraft perception under monocular imaging conditions. It contains 24,000 high-quality RGB images, each accompanied by rich ground-truth annotations including 2D semantic keypoints, 2D\3D bounding boxes, and full 6-DoF poses. The data is split into three standardized subsets—20,000 for training, 2,000 for validation, and 2,000 for testing—ensuring consistent evaluation protocols. Organized in a COCO-compatible format, the dataset includes camera intrinsic parameters and a 3D keypoint template defining all semantic landmarks in the spacecraft body frame. This resource targets three core computer vision tasks: object detection, keypoint localization, and pose estimation, providing a unified platform for developing and comparing algorithms in spaceborne applications.

Dataset Structure

Data Instance

The dataset follows a COCO-compatible format. A typical data instance contains the following fields:

• Image: RGB image containing the spacecraft.
• Annotations:
  • bbox: 2D object bounding box (normalized YOLO format).
  • bbox_3d: 3D bounding box annotations, including 3D center, dimensions, yaw, depth, projected 2D center, and projected 2D vertices.
  • keypoints_2d: 2D semantic keypoint coordinates.
  • camera_quaternion: Ground-truth camera orientation (unit quaternion).
  • camera_translation: Ground-truth camera translation (meters).
• Metadata:
  • camera_intrinsics: Camera intrinsic parameter matrix.
  • keypoint_template_3d: 3D template of all semantic keypoints in the spacecraft body frame.

Data Splits

Split Name	Number of Images
Training	20,000
Validation	2,000
Test	2,000

Supported Tasks

The SDKP dataset is designed for multi-task spacecraft visual perception and supports the following computer vision tasks:

1. 2D/3D Object Detection: Detect spacecraft in RGB images using both 2D and 3D bounding box annotations for object localization and spatial perception.

2. 2D Semantic Keypoint Detection: Predict the image coordinates of predefined semantic keypoints on the spacecraft, establishing explicit correspondences between image observations and spacecraft structures.

3. Monocular 6-DoF Pose Estimation: Estimate the six-degree-of-freedom (6-DoF) pose of the spacecraft relative to the camera from a single RGB image using the provided ground-truth pose annotations.

Annotation File Update

Update Notice:
The initial release of the SDKP dataset did not include the 3D bounding box (bbox_3d) annotations in the annotation_train.json, annotation_val.json, and annotation_test.json files.

To address this issue, we have uploaded the corrected annotation files in the Correction folder. These updated JSON files include the complete 3D bounding box (bbox_3d) annotations, while all other annotation fields and image data remain unchanged.

Users are encouraged to use the corrected annotation files in the Correction folder for all future experiments involving the SDKP dataset.