Long Video Understanding on Long Video Bench v1.0 (test)
69.84Accuracy (15s)FLoC
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| FLoCModel=Qwen2-VL-7B, Ratio=2^-3, Tokens=4320, Frames=2562025.10 | 69.84 | 72.09 | 57.04 | 50 | 57.82 | |
| FLoCModel=Qwen2-VL-7B, Ratio=2^-4, Tokens=2160, Frames=2562025.10 | 65.61 | 72.67 | — | 49.11 | 56.55 | |
| Full (No Compression)Model=Qwen2-VL-7B, Ratio=100%, Tokens=34560, Frames=2562025.10 | 64.55 | 71.51 | 54.85 | 48.05 | 55.5 | |
| FLoCModel=Qwen2-VL-7B, Ratio=2^-5, Tokens=1080, Frames=2562025.10 | 64.55 | 69.77 | 54.37 | 47.34 | 54.82 | |
| FLoCModel=Qwen2-VL-2B, Ratio=2^-5, Tokens=1080, Frames=2562025.10 | 57.67 | 60.47 | 48.06 | 40.96 | 48.02 | |
| Full (No Compression)Model=Qwen2-VL-2B, Ratio=100%, Tokens=34560, Frames=2562025.10 | 55.56 | 58.14 | 50.49 | 42.2 | 48.69 | |
| FLoCModel=Qwen2-VL-2B, Ratio=2^-4, Tokens=2160, Frames=2562025.10 | 55.56 | 61.63 | 49.03 | 43.97 | 49.44 | |
| Frame UniformModel=Qwen2-VL-2B, Ratio=2^-3, Tokens=4320, Frames=322025.10 | 53.97 | 63.95 | 47.57 | 43.79 | 48.99 | |
| PoolingModel=Qwen2-VL-2B, Ratio=2^-3, Tokens=4320, Frames=2562025.10 | 53.44 | 58.72 | 47.57 | 41.67 | 47.35 | |
| FLoCModel=Qwen2-VL-2B, Ratio=2^-3, Tokens=4320, Frames=2562025.10 | 53.44 | 60.47 | 47.57 | 44.5 | 48.77 |