thundarstrom/indian-anpr-ocr-corpus

Indian License Plate Character Recognition (PARSeq & LMDB)

Part of the Edge-AI Traffic & Vehicle Analytics System repository by thundarstrom.

Dataset Summary

A curated dataset of 18,537 normalized license plate image crops aligned strictly with Indian Motor Vehicle Act alphanumeric formats (^[A-Z]{2}[0-9]{1,2}[A-Z]{1,3}[0-9]{4}$).

Formats Provided

1. Raw Crops & Ground Truth: parseq_dataset/ with gt.txt (Tab-delimited: filename \t text).
2. High-Throughput LMDB Store: Ready for zero-bottleneck GPU memory-mapped I/O in PyTorch Lightning / PARSeq.

Splits

• Train: 15,756 crops (85%)
• Val: 1,854 crops (10%)
• Test: 927 crops (5%)
• Vocabulary: 36 alphanumeric characters (0-9, A-Z)

---

How to Access and Download

Using Automated Project Downloader (Extracts automatically)

# Clone / pull and auto-extract dataset
python scripts/download_hf_datasets.py --dataset anpr_ocr --org thundarstrom

Using huggingface_hub Python SDK

from huggingface_hub import snapshot_download

# Download into local dataset directory
local_path = snapshot_download(
    repo_id="thundarstrom/indian-anpr-ocr-corpus",
    repo_type="dataset",
    local_dir="data/datasets/parseq_dataset"
)
print(f"Dataset downloaded to: {local_path}")

---

Recommended Training / Evaluation Recipe

python scripts/train_parseq.py --data_dir data/datasets/parseq_lmdb --batch_size 256 --max_epochs 50 --lr 7e-4

---

Citation & Maintainer

• Maintained by: thundarstrom
• Project: Edge-AI Real-Time Traffic Violation Detection & ANPR
• License: CC-BY-4.0