Question about memory length

[YingYellow]

Thank you for your outstanding work. I have a question about the memory length. In Figure 5, as the length increases, the model’s accuracy first increases and then decreases. Why the model's accuracy will decrease ultimately? It seems like that more information is preserved with a larger memory.

[Espere-1119-Song]

MovieChat is a training-free method based on VideoLLaMA. Since VideoLLaMA is trained on relatively short videos, it makes it hard to understanding extremely long video for MovieChat too.

[YingYellow]

Thank you for your reply. Have you ever tried to train the VideoLLaMa on Movie-chat dataset?

[Espere-1119-Song]

Since there was no high-quality long video QA dataset last year, we didn't train VideoLLaMA