Polling-based Object Probing Evaluation (POPE) on GQA POPE Adversarial
81.92AccuracyICT w/o image
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ICT w/o imageModel=LLaVA-v1.52024.11 | 81.92 | 82.11 | |
| ICT w/o objectModel=LLaVA-v1.52024.11 | 81.65 | 82 | |
| ICT w/o objectModel=Qwen-VL2024.11 | 81.6 | 81.53 | |
| ICTModel=LLaVA-v1.52024.11 | 81.5 | 82.27 | |
| ICTModel=Qwen-VL2024.11 | 80.83 | 80.6 | |
| OPERAModel=Qwen-VL2024.11 | 80.24 | 80.64 | |
| ICT w/o imageModel=Qwen-VL2024.11 | 80.03 | 79.9 | |
| VCDModel=Qwen-VL2024.11 | 80.01 | 80.75 | |
| VCDModel=LLaVA-v1.52024.11 | 76.09 | 78.78 | |
| RegularModel=Qwen-VL2024.11 | 75.46 | 74.33 | |
| RegularModel=LLaVA-v1.52024.11 | 75.08 | 76.06 | |
| OPERAModel=LLaVA-v1.52024.11 | 75 | 78.71 |