关于复现的实验参数设置

[aub123]

您好，我在 H800 上使用脚本的默认参数测试了 Focus 方法在 Longvideobench 评估基准以及 LLaVA-video-7B 上的性能，在 61-62 间没能达到论文中的数据 63-64。
想请问一下能不能公布各个完整实验复现的参数，以便复现使用，感谢万分。

[WuSakaid]

同帖主，我在A800上使用脚本默认参数，在longvideoBench上复现结果， LLaVA-video-7B 上的性能，在 61-62 间，videomme上的结果为61.78，与论文标注结果差的有点多，想请问一下论文中的结果是使用了哪一组参数？

[chenlong50]

@ziruizhu good work，请问可以提供筛选后的关键帧json吗？

[Lucas-Song-zero]

同帖主，在RTX5090 32GB配置上，使用脚本默认参数，LLaVA-video-7B，VideoMME, 结果为61.5左右，和论文标注结果相差较大，请问论文VideoMME是否有相关更优的参数设置？

I conducted the experiment on RTX5090(32GB), with default parameters defined in the scripts of this repo, LLaVA-video-7B, resulting score of around 61.5, are there any better parameters for VideoMME?

[ziruizhu]

Thanks for your interest in FOCUS!

For the hyperparameter settings, LVB and VideoMME require distinct configurations. As discussed in the paper, queries in LVB usually contain more informative cues and are easier to match with relevant frames. In contrast, VideoMME queries are often shorter and contain less semantic information. Therefore, the keyframe selection for VideoMME should not be too aggressive compared with uniform sampling. In practice, achieving the optimal VideoMME score requires a larger zoom_ratio and final_min_arms.

As my internship has concluded, the experiment logs and server access are now maintained by the team. For the exact configurations and the selected keyframe JSON files, please feel free to email my co-author Hailun (xuhailun@tiktok.com). He led the VideoMME experiments during this project and has more direct context.