How to generate retrieval scores for a custom dataset using InternVideo2

[xeroqin]

Hi,

I would like to use InternVideo2 to obtain retrieval scores for my custom video dataset. I have a few questions:

What is the recommended way to prepare my custom dataset (video files, captions, etc.) so that InternVideo2 can process it?

Which scripts or functions should I run to generate the retrieval scores for my dataset?

Are there any specific configuration settings I need to adjust for custom datasets?

Any example workflows or best practices for generating scores outside the provided pre-extracted ones?

Thank you for your guidance!

[ikodoh]

Hi, thank you for your interest in our work.

You can refer to this annotation file, which is fully compatible for extracting the InternVideo2 score.
In addition, our retrieval_utils.py is based on the retrieval_utils.py from InternVideo2, so you may find it helpful to check that code when extracting retrieval scores.

Please let me know if you have any further questions.

Thank you!

[xeroqin]

Hi,

Thank you very much for your helpful response! I’ll go through the annotation file and retrieval_utils.py as you suggested to set up the pipeline for my dataset. This gives me a clear starting point.

If I run into further issues during the process, I’ll be sure to reach out again.
Thanks once more for your guidance and support!

Best regards!