Object Detection on In-the-Wild Construction Site Monitoring OD
87Excavators (Object Exist) IoUQwen2.5 (Baseline) (7B)
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Qwen2.5 (Baseline) (7B)Evaluation Protocol=LoRA-tuning, Global Image Scale=Full-size, Prompting Strategy=Standard2026.07 | 87 | 86.8 | 47.3 | 46 | 72.1 | 69.3 | 68.1 | 100 | |
| Qwen2.5* (7B)Evaluation Protocol=LoRA-tuning, Global Image Scale=Full-size, Prompting Strategy=Example image-based CoT2026.07 | 84.7 | 79 | 36.7 | 17.6 | 60.9 | 59.3 | 56.4 | 193 | |
| Qwen2.5 (7B)Evaluation Protocol=Zero-shot / Few-shot, Global Image Scale=Full-size2026.07 | 83.7 | 65.1 | 24.1 | 5.6 | 53.7 | 22.8 | 42.5 | 100 | |
| AVA-VLM (7B)Evaluation Protocol=LoRA-tuning, Global Image Scale=1/4 downsampled, Prompting Strategy=Adaptive Visual Attention2026.07 | 79.5 | 79.4 | 38.3 | 37.4 | 64.6 | 62.6 | 60.3 | 25.8 | |
| Grounding DINOEvaluation Protocol=Zero-shot / Few-shot, Global Image Scale=Full-size2026.07 | 71 | 45.4 | 15.3 | 2 | 11.2 | 2.1 | 24.5 | — | |
| LLaVA-v1.5 (13B)Evaluation Protocol=Zero-shot / Few-shot, Global Image Scale=Full-size2026.07 | 54.5 | 33.9 | 19 | 3.8 | 23.9 | 4.2 | 23.2 | — | |
| GPTEvaluation Protocol=Zero-shot / Few-shot, Global Image Scale=Full-size2026.07 | 35.8 | 29.3 | 18.2 | 11.6 | 10.1 | 8.6 | 18.9 | — | |
| LLaVA-v1.5 (7B)Evaluation Protocol=Zero-shot / Few-shot, Global Image Scale=Full-size2026.07 | 34.6 | 20.9 | 13.3 | 3.1 | 15 | 3.8 | 15.1 | — |