You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Vietnamese ID Document Detection Model

Abstract

This repository contains a Detectron2-based object detection model trained to localize Vietnamese identity documents in images. The detector predicts a single class, document, and is intended as a document-localization stage before downstream processing such as perspective correction, background removal, or OCR.

Model

Architecture: Faster R-CNN R50 FPN 3x
Framework: Detectron2
Detection class: document
Input: natural images containing Vietnamese identity documents
Output: bounding box coordinates with confidence scores

Training Data

The model was trained on a normalized COCO-format dataset merged from 6 local sources:

archive
behind cccd.v6i.coco
CCCD Project.v2i.coco
cccd.v2i.coco
detect cccd.v2i.coco
Project2.v2i.coco

Merged dataset summary:

Train split: 3354 images / 3347 annotations
Empty train images filtered by the trainer: 21
Effective annotated train images used for optimization: 3333
Validation split: 458 images / 452 annotations

Training Setup

Base learning rate: 0.00025
Maximum iterations: 18000
Learning-rate decay milestones: 14400, 16200
Batch size: 2
Multi-scale train resize: 640,672,704,736,768,800
Max image size: 1333

Evaluation

The final validation metrics below are from the last Detectron2 evaluation at iteration 18000.

Metric	Value
AP (bbox)	95.151
AP50 (bbox)	98.101
AP75 (bbox)	98.086
APl (bbox)	95.151

Raw training and evaluation logs are included in metrics.json.

Intended Use

This model is designed for document localization, especially as a preprocessing step before cropping or OCR. It is optimized for finding the document region, not for classifying document side, document type, or extracting text content.

Notes

This repository stores the model artifacts and model card. Training and inference scripts are maintained in the source project repository used to train the model. A TorchScript export artifact may also be included alongside the PyTorch checkpoint.

Downloads last month: -