Note · 2026-06-18
Benchmarking Qwen-VL-Max on a 200-photo crop-damage eval set
We labeled 200 photos pulled from Wikimedia Commons and Kaggle agricultural datasets across 8 crops (rice, wheat, maize, cotton, soybean, sorghum, millet, groundnut) and 6 damage causes (drought, flood, hail, frost, pest, disease).
Top-1 crop identification: Qwen-VL-Max 94%, GPT-4V 91%, Gemini 1.5 Pro 90%. Top-1 damage cause: Qwen-VL-Max 89%, GPT-4V 86%, Gemini 1.5 Pro 84%.
Severity grading (within ±10 of expert adjuster): 82% across all three models. The interesting finding wasn't who won — it's that all three are accurate enough for the human-in-the-loop workflow. The bottleneck is fraud, not perception.