Closed TicTacToePP closed 3 months ago
Thank you for your interest in my work. The memory requirements for FreeVA w/ LLaVA-1.5 are as follows:
Method | MLLM | Frames | GPU Memory | MSVD-QA Acc. |
---|---|---|---|---|
FreeVA S1 | LLaVA-1.5 7B | 4 | 16.1G | 69.6 |
FreeVA D1 | LLaVA-1.5 7B | 4 | 19.1G | 70.5 |
FreeVA D1 | LLaVA-1.5 13B | 4 | 32.0G | 71.1 |
FreeVA D2 | LLaVA-1.5 13B | 8 | 32.0G | 71.8 |
GPT-3.5-Turbo-0613 API is employed for above evaluation.
Hi, I find your paper interesting. However, as a newbie in this field, I am wondering the T (number of frames) used here. Does each dataset samples the same T which are always small? and how much GPU memory will the proposed dense aggregation consumes over prior sparse methods? Thx.