PPE mask and footwear classification - experiment notes

Experiment: PPE_Mask_Shoes_Model_Comparison.ipynb.
Models: ResNet-50, EfficientNet-B0, YOLOv8n-cls; two separate classification tasks.
Source: Personal Protective Equipment / PPEs v7 via keremberke Hugging Face mirror.
Dataset revision: 90d2c6da950a6168fcee20ec69e194a034f44eef.
The notebook retains source attribution and archive SHA-256 checksums.

Original archives are pooled because the published training split has no mask examples. Crops are grouped by source image, identical pixels, and temporal blocks. Numbered frames use 100-frame blocks with a 10-frame guard on either side of each boundary. All six runs share an assignment targeting 70/15/15 percent.

Validation macro F1 selects checkpoints and models. Recorded selection:
- Mask: ResNet-50; test accuracy 0.965517, macro F1 0.953666; 87 test crops from 3 inferred groups.
- Footwear: EfficientNet-B0; test accuracy 1.0, macro F1 1.0; 120 test crops from 4 inferred groups.

Results are from the saved Tesla T4 training run.

Limits: this is a within-recording temporal-block baseline, not an independent worker/session benchmark. Crops can share recordings or people across splits. Shoes may be detached rather than worn. Classifier confidence is uncalibrated. GPU latency excludes camera capture, detection, preprocessing, and task rules. Workshop compliance and Raspberry Pi performance are not established.

For reproducibility beyond the saved notebook, retain the exported results ZIP with weights, split manifests, predictions, run configuration, environment versions, and histories.