thundarstrom/indian-anpr-ocr-benchmark

DashCop Real-World Out-of-Distribution ANPR Benchmark

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

Dataset Summary

A frozen evaluation benchmark of 3,034 real dashcam plate crops captured in challenging conditions (low resolution median 56x34 px, motion blur, direct sunlight, rain).

Policy

• FROZEN TEST SET: Never train, augment, or fine-tune models on this dataset.
• Use exclusively to evaluate zero-shot generalization of OCR models in real Indian road conditions.

---

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_benchmark --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-benchmark",
    repo_type="dataset",
    local_dir="data/datasets/dashcop_ocr"
)
print(f"Dataset downloaded to: {local_path}")

---

Recommended Training / Evaluation Recipe

# Evaluation only\npython scripts/test_ocr.py --benchmark dashcop_ocr

---

Citation & Maintainer

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