CASTLE @ EgoVis - CVPR 2026

1st Asynchronous CASTLE Challenge at the Joint Egocentric Vision Workshop in Conjunction with CVPR 2026

We will hold a pilot task at the 3rd Joint Egocentric Vision Workshop (EgoVis) held in conjunction with CVPR 2026 in June 2026 in Denver, Colorado, USA. The challenge builds upon prior challenges at EgoVis, as well as IViSE and the 1st CASTLE Grand Challenge at ACM Multimedia 2025

Timeline

Tasks

For this initial challenge, we offer one task. The varyity of tasks will be expanded in the future.

💬 Question Answering

This first task will focus on closed-form question answering. Given the entire dataset of over 600 hours of content from 15 different perspectives, you are asked to select the correct answer to a given question out of four possible options. This tasks requires video retrieval and long-form multi-stream video understanding to be solved. The queries will be given as a single JSON document using the following structure:

{
  "2026_q0001": {
    "query": "Question formulated in English",
    "answers": {
      "a": "First Answer Option",
      "b": "Second Answer Option",
      "c": "Third Answer Option",
      "d": "Fourth Answer Option"
    }
  }, 
  "2026_q0002": {
    "query": "Another question formulated in English",
    "answers": {
      "a": "First Answer Option",
      "b": "Second Answer Option",
      "c": "Third Answer Option",
      "d": "Fourth Answer Option"
    }
  }, 
  ...
}

Submission Format

Submissions are expected as a single JSON file in the following format.

{
  "2026_q0001": "a",
  "2026_q0002": "c",
  ...
}

Submissions can be made via Codabench (link coming soon).