kv1388/FANVID-Face_and_License_Plate_Recognition_in_Low-Resolution_Videos

FANVID: Face and License Plate Recognition in Low-Resolution Videos

Overview

FANVID is a benchmark dataset designed to advance research in face detection and matching and license plate recognition under challenging low-resolution surveillance video conditions. Unlike existing datasets, FANVID features faces and license plates that are unrecognizable in individual frames, encouraging models to leverage temporal context across video sequences for improved recognition.

---

Dataset Composition

• Face Detection and Matching

  • Videos with low-resolution faces.
  • Task: Detect faces in video clips and match to high-resolution gallery mugshots.
  • Contains distractor faces to simulate real-world surveillance challenges.

• License Plate Recognition

  • Vehicle videos with low-resolution license plates.
  • Task: Detect plates and transcribe text without prior knowledge of plate numbers.
  • Includes distractor license plates.

---

Annotations

• Bounding boxes per frame for faces and license plates.
• Frame-level indexing.
• Identity labels for faces and ground-truth text for license plates.
• All annotations are manually verified on high-resolution source videos, then downsampled to ensure targets are not clearly visible in any single frame.

---

Evaluation Metrics

• Intersection-over-Union (IoU): Measures detection accuracy for bounding boxes.
• Edit Distance: Evaluates OCR accuracy on license plate text.
• Metrics are designed to prioritize accurate recognition while tolerating minor bounding box misalignments.

---

🚀 Quickstart

You can programmatically download and set up the FANVID dataset using the Hugging Face Hub:

from huggingface_hub import snapshot_download
import shutil

# Step 1: Download the dataset repository
repo_dir = snapshot_download(
    repo_id="kv1388/FANVID-Face_and_License_Plate_Recognition_in_Low-Resolution_Videos",
    repo_type="dataset"
)

# Step 2: Zip it (optional)
shutil.make_archive("FANVID_dataset", 'zip', repo_dir)

Then, from your terminal or notebook:

# Step 3: Unzip the dataset
unzip -q FANVID_dataset.zip -d FANVID_dataset

# Step 4: List extracted contents
ls FANVID_dataset

# Step 5: Install required tools
pip install yt_dlp

# Step 6: Download videos and metadata
python FANVID_dataset/assets/dataset_script_celebs.py
python FANVID_dataset/assets/download_script_lp.py

If there are any failed downloads logged, re-run:

python FANVID_dataset/assets/failed_LP_download.py

You can now begin using the dataset and verifying annotations:

• Annotation files are stored in the FANVID_dataset/data/ folder.
• Visualize and inspect using the Jupyter notebook:

jupyter notebook FANVID_dataset/assets/FaceTextRecBox.ipynb

---

Usage

• Develop and benchmark video-based face and license plate recognition algorithms that utilize temporal context.
• Test super-resolution, detection, and recognition models under real-world surveillance conditions.

---

Baselines

We provide baseline results using a pipeline combining state-of-the-art video super-resolution, detection, and recognition models. Code for evaluation and baseline implementation will be released shortly alongside the dataset.

---

License and Citation

• License: [CC BY 4.0]
• Citation:

If you use this dataset, please cite:

@misc{viswanathan2025fanvidbenchmarkfacelicense,
  title={FANVID: A Benchmark for Face and License Plate Recognition in Low-Resolution Videos},
  author={Kavitha Viswanathan and Vrinda Goel and Shlesh Gholap and Devayan Ghosh and Madhav Gupta and Dhruvi Ganatra and Sanket Potdar and Amit Sethi},
  year={2025},
  eprint={2506.07304},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2506.07304}
}

---

Contact

For questions or collaboration, please contact: [[email protected]]