Question: Application to LLaVA-OV and Qwen2vl

[arsoyul]

Thank you for sharing your great work!

I want to experiment on the LLaVA-OV and Qwen2VL. If you have a similarity threshold on this model, can you share me? I found Qwen2VL's threshold is 0.5 in your example code. Is it correct?
Also i want to ask about how to select the similarity threshold on other models.
In your paper,

Since Similarity distributions vary across models, we set Sthreshold to match the median of similarity at the first model layer under typical input cases, such as 128 samples from the VideoMME dataset.

Then, the mean of median similarity across all VideoMME samples is the similarity threshold?

[M-X-05]

Hi @arsoyul, sorry for the delayed response. For Qwen2VL, the similarity threshold of 0.5 in the example code is correct. For other models (e.g., LLaVA-OV), the threshold can be determined based on their similarity distribution. Following the paper, you can sample 128 videos from VideoMME, compute the median similarity at the first layer for each sample, and take the mean of these medians as the threshold. Using all videos from VideoMME is also fine if you want a more stable estimation.