Vehicle Color Recognition on UFPR-VeSV
94.6Mi-AccProposed approach
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Proposed approachTraining Data=Gemini-augmented, Loss=WCE, Scheduler=LWCD, Augmentation=color-safe augmentation, Preprocessing=foreground-aware preprocessing, Fusion=hard-voting ensemble fusion2026.06 | 94.6 | 79.7 | 78.8 | |
| DINOv3-LargeLoss=CE, Scheduler=Cosine Decay (CD), Data Augmentation=None2026.06 | 94 | 72.5 | 74.4 | |
| EfficientNet-V2Source=Lima et al. [28]2026.06 | 93.5 | 71.5 | 73.8 | |
| Swin-TLoss=CE, Scheduler=Cosine Decay (CD), Data Augmentation=None2026.06 | 93.2 | 72.5 | 74 | |
| DINOv3-BaseLoss=CE, Scheduler=Cosine Decay (CD), Data Augmentation=None2026.06 | 93 | 69.6 | 72.3 | |
| ViT-B/16Loss=CE, Scheduler=Cosine Decay (CD), Data Augmentation=None2026.06 | 92.8 | 69.9 | 72.5 | |
| DINOv3-SmallLoss=CE, Scheduler=Cosine Decay (CD), Data Augmentation=None2026.06 | 92.5 | 69.1 | 71.2 | |
| ResNet-50Loss=CE, Scheduler=Cosine Decay (CD), Data Augmentation=None2026.06 | 91.2 | 64.1 | 66.4 | |
| EfficientNet-V2Loss=CE, Scheduler=Cosine Decay (CD), Data Augmentation=None2026.06 | 90.9 | 66.7 | 68.2 |