The MedVidU Challenge runs in two phases. Phase 1 is a public leaderboard on the MedVidBench test set. The top 5 teams from Phase 1 advance to Phase 2, where their models are evaluated on a completely held-out test set via a provided Docker image. Follow the steps below.

Phase 1 — Public Leaderboard

1

Download the training data and train your model

  • Download the MedVidU ECCV 2026 train/val split from UII-AI/MedVidU_ECCV2026_TrainVal on Hugging Face.
  • Sign in to Hugging Face, click Request access, fill out the short form, and agree to the terms of use. Requests are reviewed manually and typically approved within a few business days.
  • Once approved, you may use this split for training and local validation, alongside any other open-source or internal data of your choice.
2

Submit predictions to the MedVidBench Leaderboard

Run inference on the MedVidBench test set and submit your predictions to the MedVidBench Leaderboard.

Include the following with your submission:

  • Model Name
  • Team Name
  • Authors
  • Institution
  • Email ID — mandatory for MedVidU ECCV 2026 Challenge participants
  • Remark: ECCV challenge — required to flag your entry as part of the challenge
3

Submit a report on OpenReview Required

All challenge participants are required to submit a report describing their method on OpenReview for the MedVidU Challenge. The submission link will be announced soon.

Phase 2 — Held-Out Evaluation (Top 5 Teams) (starts on July 13)

4

Download the evaluation Docker image Coming soon

The top 5 teams on the Phase 1 leaderboard will be invited to Phase 2. We will release a base Docker image with the evaluation harness and required entry-point interface. Build your submission on top of this image so that it is compatible with our internal evaluation pipeline.

6

Submit your Docker image

Package your trained model inside the provided Docker image and submit it to the organizers by email. We will run your container on the completely held-out test set to determine the final challenge ranking.

Confidentiality. We encourage teams to open-source their models to benefit the community, but this is not required. Teams that wish to keep their models private can do so without concern: submitted Docker images (and the model weights they contain) are treated as confidential. They will be accessed only by the organizing team for the sole purpose of running the held-out evaluation, will not be shared with any third party, will not be redistributed or open-sourced by us, and will be permanently deleted from our infrastructure once final rankings are determined.