Multi-party Multi-modal Social Interaction (MMSI) on YouTube
64.9STI AccuracyOnline-MMSI-VLM (LLaMA)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Online-MMSI-VLM (LLaMA)Backbone=LLaMA2025.03 | 64.9 | 68.5 | 48.4 | 60.6 | |
| Online-MMSI-VLM (Qwen)Backbone=Qwen2025.03 | 64.6 | 68.8 | 47.2 | 60.2 | |
| Vanilla LLaMABackbone=LLaMA2025.03 | 63.2 | 65.4 | 46.6 | 58.4 | |
| Vanilla InternVLBackbone=InternVL2025.03 | 62.4 | 65.1 | 47 | 58.2 | |
| Vanilla QwenBackbone=Qwen2025.03 | 60.8 | 62.9 | 45.8 | 56.5 | |
| Lee et al. (2024a) (Offline)Temporal Context=Offline2025.03 | 60 | 63.5 | 43.2 | 55.6 | |
| Lee et al. (2024a) (Online)Temporal Context=Online2025.03 | 59.1 | 63.4 | 47.3 | 56.6 | |
| GPT-4o (Zero-shot)Evaluation Protocol=Zero-shot2025.03 | 38.5 | 35.4 | 37.9 | 37.3 | |
| Gemini 2.5 Pro (Zero-shot)Evaluation Protocol=Zero-shot2025.03 | 23.3 | 26.7 | 16.7 | 22.3 |