| license: apache-2.0 | |
| library_name: anomalib | |
| pipeline_tag: image-classification | |
| tags: | |
| - anomaly-detection | |
| - computer-vision | |
| - pcb | |
| - efficientad | |
| - robustad | |
| # VisionQC EfficientAD-medium for RobustAD PCB | |
| EfficientAD-medium checkpoint trained on the RobustAD PCB source domain with | |
| Anomalib. The best checkpoint was selected using image AUROC on a 20% | |
| synthetic validation split with early stopping. | |
| ## Benchmark | |
| | Domain | Image AUROC | Pixel AUROC | Pixel AUPRO | | |
| | ------ | ----------: | ----------: | ----------: | | |
| | test0 | 0.8221 | 0.9237 | 0.6809 | | |
| | test1 | 0.5000 | 0.6878 | 0.2052 | | |
| | test2 | 0.5000 | 0.7277 | 0.2539 | | |
| | test3 | 0.4917 | 0.8338 | 0.5107 | | |
| | test4 | 0.5000 | 0.8205 | 0.5082 | | |
| | test5 | 0.6250 | 0.8739 | 0.6213 | | |
| The checkpoint is a baseline: image-level performance degrades substantially | |
| under domain shift. See `domain_metrics.csv`, `ard.csv`, and `benchmark.json` | |
| for the complete evaluation output. | |
| ## Download | |
| ```bash | |
| hf download thangkt/visionqc-efficientad-medium-robustad-pcb \ | |
| model-best.ckpt --local-dir weights/robustad-pcb | |
| ``` | |
| Project: <https://github.com/pnthang04/VisionQC> | |