Dataset Description

This dataset comprises 1,046 real grayscale images acquired by sampling 22 rectilinear trajectories under systematically varied illumination conditions. In addition, 40,000 synthetic grayscale images were generated and used for training purposes.

The complete dataset, together with the trained model weights, is provided as a compressed archive
(**data_weights.tar.gz, approximately 3.9 GB)

Archive Contents

Data

• data/real/images/
  1,046 real grayscale test images
• data/real/labels_real.json
  Ground truth annotations, including keypoints, bounding boxes, and pose information
• data/synthetic/images/
  40,000 synthetic training images
• data/synthetic/split_yolo/
  YOLO-format dataset split
  (35,000 training / 5,000 validation / 772 test images)
• data/synthetic/gan_synth2real_images/
  Synthetic images translated to the real domain using CycleGAN
• data/cam_sat.json
  Camera intrinsic parameters and spacecraft model

Model Weights

• weights/odn/
  YOLO11n object detection network weights provided in PyTorch, ONNX, and TensorRT formats (FP16 and FP32)
• weights/krn/
  EfficientNet-B0 keypoint regression network weights provided in PyTorch, ONNX, and TensorRT formats (FP16, FP32, and INT8)

Data Acquisition and Characteristics

For the real dataset, the target-to-camera distance ranged from 0.33 m to 0.79 m, measured from the geometric center of the physical model. For the synthetic dataset, distances ranged from 0.3 m to 1.0 m.

All images were converted to grayscale to improve computational efficiency and to accelerate processing within the developed pose estimation pipeline.

Illumination conditions were systematically varied for each test case with respect to:

• the α and β illumination angles,
• the position of the sun simulator along the z and y axes, and
• the intensity of the simulated sunlight.

Extraction Instructions

tar -xvzf data_weights.tar.gz -C /path/to/destination/