Welcome to the first instance of the iMED Challenge a subset of the EndoVis Challenge.

The official Task 1 dataset has now been posted. Please see the GitHub example baseline: https://github.com/imed-challenge/imedpe

Note for early downloaders:
If you downloaded Task 1 during the first week it was available, please do not use that version. We accidentally released the wrong subset, and the pose.txt file was not aligned with the frame numbers. The sequences have now been reshuffled and the poses corrected. Using the old poses will very likely result in worse performance.

Update regarding key dates:
Because the Task 1 dataset and Docker setup instructions were only uploaded recently, we will be extending the public leaderboard submission from May 30th to June 30th. This is to give participants additional time to develop and test their solutions.

We have also posted Docker setup instructions for each task in the challenge files and specific instructions for building the dockers for each subtask in the baseline example repositories.

Task-Specific Notes

For Task 2, because the baseline requires custom CUDA kernels, we released a starting repository. However, participants are welcome to use any suitable method — Gaussian splatting is not the only possible solution to this problem. For example: https://onlinelibrary.wiley.com/doi/full/10.1111/cgf.70494

For Task 1, participants may consider pose estimation models, including:

• Domain-specific approaches: https://pmc.ncbi.nlm.nih.gov/articles/PMC10329609/
• General feature detection and matching algorithms: https://link.springer.com/chapter/10.1007/978-3-031-91989-3_8

It may also be interesting to use the stereo images.

Rules and Registration

Please make sure you have emailed one of the organizers, Sierra Bonilla ( sierra.bonilla.21@ucl.ac.uk) with a copy of the signed EndoVis Rules. Please include [iMED Challenge 2026] in the email subject line.

Frequently Asked Questions

Task 1

Metrics:
We are partnered with the CliMB sub-challenge, so we use their metric scripts. However, note that we are comparing cam2 to cam1 poses, which is not the same as relative trajectory estimation on a single camera. Please be careful with pose conventions when using other models or training your own.

Metric script: https://github.com/imed-challenge/imedpe/blob/main/scripts/evaluate_ate.py

Task 2

Tool Occlusion:
Will tools be excluded from the final evaluation metrics?
Yes. We provide tool masks and load them in the metric scripts.

Field of View (FOV):
Since the two endoscopes have non-overlapping regions, how should these areas be handled during evaluation?
Please see the metric script: https://github.com/smbonilla/Endo-4DGS/blob/main/metrics.py
We avoid evaluating non-overlapping regions.

Will evaluation be based entirely on performing inference on Endoscope 1 using information from Endoscope 2?
Yes. Evaluation will be performed entirely on the unseen Endoscope 1. Please do not load Endoscope 1 frames during training.

From now on, please post questions in the Discussions section below so all participants can see the answers.

Best,
Sierra Bonilla, John Han, and Tianyi Song