Thanks for releasing this amazing work.
As paper described, the current VLM-based work suffer from lower fps due to their larger size.
Since the fps of carllava and your work are quite speicific in the paper, i wonder that do you have some script for measuring the fps for VLM ?
If it's not contained in your code base, could you give me some suggestion that how should i measure fps, which can produced the fps results consist with the paper description ?
2B model : 300 ms ~ 3.3 fps ; carllava model : 102 ms ~ 9.8 fps
Thanks for releasing this amazing work.
As paper described, the current VLM-based work suffer from lower fps due to their larger size.
Since the fps of carllava and your work are quite speicific in the paper, i wonder that do you have some script for measuring the fps for VLM ?
If it's not contained in your code base, could you give me some suggestion that how should i measure fps, which can produced the fps results consist with the paper description ?