Polling-based Object Probing Evaluation (POPE) on GQA POPE (Random)
89.6AccuracyICT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ICTModel=LLaVA-v1.52024.11 | 89.6 | 89.44 | |
| ICT w/o imageModel=LLaVA-v1.52024.11 | 89.03 | 89.99 | |
| ICT w/o objectModel=LLaVA-v1.52024.11 | 88.97 | 89.3 | |
| OPERAModel=LLaVA-v1.52024.11 | 88.13 | 88.91 | |
| ICT w/o objectModel=Qwen-VL2024.11 | 87.2 | 86.8 | |
| VCDModel=LLaVA-v1.52024.11 | 86.65 | 86.99 | |
| ICTModel=Qwen-VL2024.11 | 86.38 | 86.96 | |
| OPERAModel=Qwen-VL2024.11 | 86.02 | 85.29 | |
| VCDModel=Qwen-VL2024.11 | 85.59 | 85.33 | |
| ICT w/o imageModel=Qwen-VL2024.11 | 85.2 | 85.9 | |
| RegularModel=LLaVA-v1.52024.11 | 83.73 | 82.85 | |
| RegularModel=Qwen-VL2024.11 | 80.97 | 79.01 |