Context
In a recent work, we developed a real-time pipeline for pose estimation of spacecraft from single images. A key contribution was a new mathematical formulation that significantly improved accuracy and led to a WACV 2026 publication. A summary of the paper can be found under: https://pierreancey.org/projects/fastpose-vit.
Following this success, we want to expand this work in two directions:
- Strengthening the mathematical foundation, as it is currently only heuristic based.
- Testing whether the improvement generalizes to standard pose estimation datasets, rather than only space-related ones.
This proposal concerns the second direction.
Objective
The goal of this project is to adapt our existing spacecraft pose estimation pipeline so that it can run on standard benchmarks, particularly the BOP benchmark.
BOP is a widely used collection of pose estimation datasets (e.g. YCB-V, TLESS, LM-O), all sharing a common data format and evaluation protocol. Unlike our spacecraft images, these datasets often contain several objects per image, while our current pipeline only supports a single object.
The task is to make the pipeline compatible with at least one BOP dataset and run it end-to-end. The code repository will be provided.
Student contribution
Student Tasks
- Integrate at least one BOP dataset into the existing pipeline
- Update the dataloader to support multiple objects per image
- Ensure the pipeline can run end-to-end with BOP-format inputs
- (Optional) Enable submission to the BOP evaluation server
Success Criteria
A successful outcome is full support for one BOP dataset.
A stronger outcome is support for several BOP datasets.
If results are promising, this could provide a path to a publication at a larger conference.
Requirements
- Experience with PyTorch
- Comfortable working with datasets and dataloaders
Contact
If this project interests you, please send an email to Pierre Ancey ([email protected])