File size: 2,259 Bytes
68df23d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
---
license: other
license_name: educational-use-only
license_link: https://github.com/hongming111/SNAIC_Week4_SmartCart/blob/main/LICENSE
tags:
- computer-vision
- object-detection
- image-classification
- gradio
- yolo
- dinov2
- onnx
---

# 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](https://github.com/hongming111/SNAIC_Week4_SmartCart).

Powers the companion Space: [Hongming111/SNAIC_Week4_SmartCart](https://huggingface.co/spaces/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.