| # ParcelGuard Model Card |
|
|
| ## Model purpose |
| Experimental reference-guided anomaly scoring for product images. |
|
|
| ## Inputs |
| - One query product image. |
| - One or more reference images for the same item. |
|
|
| ## Outputs |
| - P95 anomaly score. |
| - Binary label from a provisional fixed threshold. |
| - Spatial anomaly map and overlay. |
|
|
| ## Architecture |
| ResNet18 pretrained on ImageNet, using layer2 spatial features. |
| Cosine distances are computed at corresponding feature-map locations; |
| the minimum distance across references is used at each location. |
|
|
| ## Evaluation |
| Current sample: 100 queries, 37 defective and 63 normal. |
| Exploratory classifier cross-validation: balanced accuracy 0.594, ROC-AUC 0.607. |
| Exploratory P95-only ROC-AUC: approximately 0.684. |
| These are exploratory results, not a held-out benchmark. |
|
|
| ## Failure example |
| One defective sample had a P95 score of approximately 0.654 and was classified |
| as NORMAL using the provisional threshold of 0.70. |
|
|
| ## Limitations and risks |
| The method is sensitive to view changes, alignment, packaging, texture and lighting. |
| The heatmap is not verified pixel-level defect localization. |
| The threshold was selected on the exploratory sample and requires independent validation. |
|
|
| ## Intended use |
| Research and experimentation only. Do not use as an autonomous quality-control |
| decision system without rigorous validation on representative held-out data. |
|
|
| ## Data and licensing |
| The project uses the Amazon Kaputt dataset. Review its CC BY-NC-ND 4.0 terms. |
| No original dataset images are intentionally included in this repository. |
|
|
