SmartCart SG — trained bundle (SNAIC Week 4 Capstone)
Trained artifacts for a self-checkout computer-vision pipeline (detect -> recognize -> price -> receipt) built on a home-grown Singapore grocery dataset. Source code and full writeup: hongming111/SNAIC_Week4_SmartCart.
Powers the companion Space: Hongming111/SNAIC_Week4_SmartCart.
Contents
| File | Purpose |
|---|---|
detector.pt, detector_v0.pt |
YOLO11 product detector (basket-hardened / earlier version) |
FastSAM-s.pt |
FastSAM "segment everything" proposals for Basket mode |
head.pt, head.onnx, head.int8.onnx |
25-way recognition head on frozen DINOv2 features (PyTorch / ONNX / quantized) |
gallery_index.npy, gallery_meta.csv |
DINOv2 embedding gallery used for open-set gating |
catalog_prices.csv |
Real median FairPrice SGD prices per class |
confidence_threshold.json |
Calibrated serving threshold (0.6, held-out crop split) |
labels.csv, manifest.json, crops/, crops_manifest.csv, crops_train.csv |
Training/label provenance |
decisive_lift_table.csv, lift_table.csv, lift_table_vs_real_photo.csv, per_class_metrics.csv, error_report.md |
Evaluation and augmentation-lift measurements |
Headline results
- Detection: YOLO11n, mAP50-95 0.76; basket-hardened retrain: real 5-product basket 0 -> 7 boxes.
- Recognition: 25-way linear head on frozen DINOv2, 94% validation (crop-aware).
- Serving (
head.onnx/head.int8.onnx): 76.5% on 17 held-out real photos, 96.2% on held-out detector crops; ONNX parity 1.4e-6; int8 = 10.6 KB.
License / usage restriction
Educational use only. Product studio images/prices originate from FairPrice/NTUC, collected under fair-dealing/educational use for coursework — do not use this repo or its contents commercially or redistribute the dataset. See the linked LICENSE for full terms.