
摘要
近期在大规模语言模型(Large Language Models, LLM)方面的进展推动了图像-语言对话代理的多种进步,但如何构建一个高效的基于视频的对话系统仍处于探索阶段。考虑到大规模语言模型和视觉主干网络的庞大体量,留给有效时序建模的GPU内存非常有限,而时序建模对于理解和回应视频内容至关重要。为此,我们提出了一种新的方法——分支时序适配器(Branching Temporal Adapter, BT-Adapter),用于将图像-语言预训练模型扩展到视频领域。具体而言,BT-Adapter作为预训练视觉编码器旁的一个即插即用的时序建模分支,在保持主干网络冻结的情况下进行微调。只需一次预训练,BT-Adapter即可无缝集成到所有使用此版本CLIP的图像对话模型中,实现无需视频指令的视频对话功能。此外,我们在分支内部开发了一种独特的非对称标记掩码策略,并为BT-Adapter设计了定制化的训练任务,从而加速收敛并获得更好的结果。得益于BT-Adapter,我们能够在不增加过多GPU成本的情况下增强现有多模态对话模型的视频理解能力。无需额外复杂的配置,BT-Adapter实现了以下几点:(1) 在各种视频任务上以较少的GPU小时数达到了最先进的零样本性能;(2) 无需任何视频指令微调的情况下优于当前的视频聊天机器人;(3) 经过视频指令微调后,在视频聊天方面取得了远超以往最佳水平的结果。
代码仓库
farewellthree/BT-Adapter
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| vcgbench-diverse-on-videoinstruct | BT-Adapter | Consistency: 2.27 Contextual Understanding: 2.59 Correctness of Information: 2.20 Dense Captioning: 1.03 Detail Orientation: 2.62 Reasoning: 3.62 Spatial Understanding: 2.35 Temporal Understanding: 1.29 mean: 2.19 |
| video-based-generative-performance | BT-Adapter (zero-shot) | Consistency: 2.2 Contextual Understanding: 2.89 Correctness of Information: 2.16 Detail Orientation: 2.46 Temporal Understanding: 2.13 mean: 2.46 |
| video-based-generative-performance | BT-Adapter | Consistency: 2.46 Contextual Understanding: 3.27 Correctness of Information: 2.68 Detail Orientation: 2.69 Temporal Understanding: 2.34 mean: 2.69 |
| video-based-generative-performance-1 | BT-Adapter | gpt-score: 2.68 |
| video-based-generative-performance-1 | BT-Adapter (zero-shot) | gpt-score: 2.16 |
| video-based-generative-performance-2 | BT-Adapter | gpt-score: 2.46 |
| video-based-generative-performance-2 | BT-Adapter (zero-shot) | gpt-score: 2.2 |
| video-based-generative-performance-3 | BT-Adapter | gpt-score: 3.27 |
| video-based-generative-performance-3 | BT-Adapter (zero-shot) | gpt-score: 2.89 |
| video-based-generative-performance-4 | BT-Adapter (zero-shot) | gpt-score: 2.46 |
| video-based-generative-performance-4 | BT-Adapter | gpt-score: 2.69 |
| video-based-generative-performance-5 | BT-Adapter (zero-shot) | gpt-score: 2.13 |
| video-based-generative-performance-5 | BT-Adapter | gpt-score: 2.34 |
| video-question-answering-on-activitynet-qa | BT-Adapter (zero-shot) | Accuracy: 46.1 Confidence score: 3.6 |
| zero-shot-video-retrieval-on-activitynet | BT-Adapter | text-to-video R@1: 37.0 text-to-video R@10: 78.9 text-to-video R@5: 66.7 |
| zero-shot-video-retrieval-on-didemo | BT-Adapter | text-to-video R@1: 35.6 text-to-video R@10: 72.6 text-to-video R@5: 61.9 |
| zero-shot-video-retrieval-on-lsmdc | BT-Adapter | text-to-video R@1: 19.5 text-to-video R@10: 45.0 text-to-video R@5: 35.9 |
| zero-shot-video-retrieval-on-msr-vtt | BT-Adapter | text-to-video R@1: 40.9 text-to-video R@10: 73.5 text-to-video R@5: 64.7 |
| zeroshot-video-question-answer-on-activitynet | BT-Adapter (zero-shot) | Accuracy: 46.1 Confidence Score: 3.2 |
| zeroshot-video-question-answer-on-msrvtt-qa | BT-Adapter (zero-shot) | Accuracy: 51.2 Confidence Score: 2.9 |
| zeroshot-video-question-answer-on-msrvtt-qa | BT-Adapter (zero-shot) | Accuracy: 51.2 Confidence Score: 2.9 |
| zeroshot-video-question-answer-on-msvd-qa | BT-Adapter (zero-shot) | Accuracy: 67.0 Confidence Score: 3.6 |
| zeroshot-video-question-answer-on-msvd-qa | BT-Adapter (zero-shot) | Accuracy: 67.0 Confidence Score: 3.6 |