Multimodal Dataset of Space Debris

Subsequently, the 3D database of space debris was built combined with 3D point cloud reconstruction technology. First, a 3D scanner was adopted to acquire submillimeter-level morphology and feature data. Next, 3D models were reconstructed via point cloud fusion and mesh generation techniques, followed by optimization processes including outlier and overlap removal, surface smoothing, mesh simplification and hole repair.

Precise 3D models were finally obtained to form the complete 3D dataset.# 2. Visible Light Dataset PLY models were imported into Blender. An image acquisition method was applied to build a multi-angle visible light dataset by adjusting the rotation angle of models.# 3.

Infrared Dataset Taking the DR-AVIT model as the core reference and the DroneVehicle aerial visible-infrared dataset as the foundation, cross-modal image translation between visible light and infrared spectra was realized. High-quality infrared datasets of space debris with diverse radiation characteristics and background interferences were thereby generated.# 4. SAR Dataset Cycle-Consistent Generative Adversarial Network (CycleGAN) was adopted to implement cross-domain translation between visible light and Synthetic Aperture Radar (SAR), converting visible light images into SAR images.# 5.

Dataset Scale The dataset adopted in this paper covers three modalities: visible light, infrared and SAR. Each modality contains 29 categories of space debris, namely momentum wheel, momentum wheel fragments, Satellite 1, Satellite 2, unassembled circuit board, Battery 1, Battery 2, battery slot, fragmented momentum wheel, fragmented solar panel, fragmented Battery 1, fragmented Battery 2, fragmented circuit board, fragmented lens, white lens, Lens Fragments 1–12, Black Lens 1 and Black Lens 2.One image was captured every 15° of rotation for each 3D model, producing 24 × 24 = 576 original images per category.

Data augmentation was performed in this paper by capturing one image every 30° of rotation, which yielded 576 × 12 = 6912 augmented images for each category. With 29 categories in total, each modality contained 6912 × 29 = 200448 augmented images. To reduce upload volume, only original images (576 images per category) are uploaded; users may perform self-defined data augmentation as required.# 6.

Project

Support This work was supported by the National Science and Technology Major Project (2022ZD0117301), the National Key Basic Research Project (2022JCJQZD20600), the Key Support For Top Innovation Team Projects (No. 2022-901), the Hundred Talents Program for Scientists (

Project

No. 2022-500), the National Excellent Professional and Technical Talents (Under 55,

Project

No. 2022-641). This dataset is used to supplement the publication of the paper "A Space Debris Recognition

Method

Based on Multimodal Feature Fusion" (DOI: 10.3724/ati2026023).