ebowwa
Normalize: All IDs now 0-23 (COCO + metadata.jsonl unified)
e5eef74
metadata
license: mit
task_categories:
  - object-detection
language:
  - en
tags:
  - currency
  - USD
  - money-detection
  - COCO
  - RF-DETR
  - YOLO
size_categories:
  - 1K<n<10K

USD Side Detection Dataset (Front/Back)

A refined COCO-format dataset for detecting US Dollar currency and classifying whether the front or back side is visible.

Dataset Summary

  • Total Images: 3,618
  • Total Annotations: 3,746
  • Format: COCO + HuggingFace JSONL
  • Classes: 24 (denominations × front/back × authentic/counterfeit)
  • Classification Accuracy: 100% (all Front/Back classified)
Split Images Annotations
Train 2,671 2,738
Valid 597 627
Test 350 381

Class Mapping (24 classes)

ID Class ID Class
0 100USD-Back 12 Counterfeit 100 USD Back
1 100USD-Front 13 Counterfeit 100 USD Front
2 10USD-Back 14 Counterfeit 10USD Back
3 10USD-Front 15 Counterfeit 10USD Front
4 1USD-Back 16 Counterfeit 1USD Back
5 1USD-Front 17 Counterfeit 1USD Front
6 20USD-Back 18 Counterfeit 20USD Back
7 20USD-Front 19 Counterfeit 20USD Front
8 50USD-Back 20 Counterfeit 50USD Back
9 50USD-Front 21 Counterfeit 50USD Front
10 5USD-Back 22 Counterfeit 5USD Back
11 5USD-Front 23 Counterfeit 5USD Front

Breakdown:

  • 12 Regular USD: Front/Back for $1, $5, $10, $20, $50, $100
  • 12 Counterfeit USD: Front/Back for $1, $5, $10, $20, $50, $100

Note: All $2 bills and generic annotations removed - only Front/Back classified data remains.

Annotation Refinement

This dataset was refined using Roboflow's usd-classification/1 model:

Phase 1: Regular USD ✅

  • Reclassified 2,236 generic labels to Front/Back variants
  • 97% success rate

Phase 2: Counterfeit USD ✅

  • Reclassified 943 counterfeit annotations across all splits (train/valid/test)
  • 97.8% success rate (269/275 in valid/test, 674/762 in train)
  • Only 13 annotations remain generic (SSL errors during classification)

Phase 3: Data Cleaning ✅

  • Removed 289 $2 bill annotations (146 regular + 143 counterfeit)
  • Reason: Model lacks "two-front"/"two-back" classes, generalization only 75% accurate

Final Statistics

  • 3,746 annotations - 100% classified to Front/Back
  • 24 classes - 12 regular + 12 counterfeit
  • 0 $2 bills (all 289 removed - 146 regular + 143 counterfeit)

Usage

from datasets import load_dataset

dataset = load_dataset("ebowwa/usd-side-coco-annotations")

Or download directly and extract for use with YOLO/RF-DETR training.

Source

Original dataset from Roboflow - "Front/Back of USD 2" project.

Refined using automated Roboflow classification API with incremental saving for fault tolerance.

License

MIT