Visual Question Answering on PolarVQA (test)
7.57Glass Count ScorePolarVLM
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| PolarVLMArchitecture=End-to-End Multimodal Pipeline2026.05 | 7.57 | 5.51 | 6.87 | 5.85 | 4.09 | 6.08 | |
| Glass Seg. + GPT-4.1Architecture=Cascaded Polarimetric Pipeline2026.05 | 7.16 | 5.48 | 6.54 | — | — | — | |
| LLaVA-1.5-13BType=Controlled Baseline, Modality=RGB-only2026.05 | 5.65 | 4.88 | 5.77 | 4.62 | 3.23 | 4.85 | |
| Qwen2.5-VL-7BType=Open-Weight RGB-only VLM2026.05 | 5.22 | 5.38 | 6.65 | 5.54 | 3.1 | 5.42 | |
| GPT-4.1Type=Proprietary VLM2026.05 | 4.81 | 5.46 | 5.96 | 5.66 | 3.71 | 5.38 | |
| InternVL3-14BType=Open-Weight RGB-only VLM2026.05 | 4.55 | 5.4 | 6.55 | 5.02 | 2.81 | 5.01 | |
| Glass Seg. + LLaVA-1.5-13BArchitecture=Cascaded Polarimetric Pipeline2026.05 | 2.94 | 4.29 | 4.54 | — | — | — | |
| BLIP2-Flan-T5-XLType=Open-Weight RGB-only VLM2026.05 | 2.53 | 3.55 | 3.05 | 3.03 | 2.06 | 2.94 | |
| PolarFree + GPT-4.1Architecture=Cascaded Polarimetric Pipeline2026.05 | — | — | — | 5.71 | — | — | |
| PolarFree + LLaVA-1.5-13BArchitecture=Cascaded Polarimetric Pipeline2026.05 | — | — | — | 4.34 | — | — |