Monocular Spacecraft Pose Estimation with Vision Transformers

In this work, we modified Mobile-URSONet backbone to create a low-weight ViT-based spacecraft pose estimator suitable for quick inference on a real spacecraft, while increasing position estimation over the baseline. Furthermore, we developed alternative versions with RepViT and hybrid Tiny-ViT/Mobile-Net backbones, improving the orientation estimation over the original model.

Instructions:

• Create a folder named datasets into the folder code
• Download the speed dataset from https://zenodo.org/records/6327547 and put it in the folder datasets. You can also run inference on speed+ lab data by downloading the speed+ dataset from https://purl.stanford.edu/wv398fc4383 and recreating the same folder structure of speed
• Modify code/src/config.py with the desired configurations, including the desired model and either training or evaluation mode. The pre-trained models are available in the code/models folder
• from the folder code/src run python main.py

Blog Post and Storyline:

• You can find a blog post with a detailed description of the project's aim, experiments, and results at https://oasis-soul-ee5.notion.site/Monocular-Spacecraft-Pose-Estimation-with-Vision-Transformers-60d53328311a4b7482b2d4903b322126
• If you prefer a quicker overview of the work, you can find the project storyline in this repo as a pdf file

This work is an extension of the following paper:

Posso, J., Bois, G., & Savaria, Y. (2022). Mobile-URSONet: an Embeddable Neural Network for Onboard Spacecraft Pose Estimation. In 2022 IEEE International Symposium on Circuits and Systems (ISCAS) (pp. 794-798). https://doi.org/10.1109/ISCAS48785.2022.9937721