Filography: A Revisited Technique for Digital Imagery

Oumar Ibrahim and Mohamad Khairy
Department of Computer Engineering, University of Sharjah, Sharjah, UAE
u22200741@sharjah.ac.ae | mishak@sharjah.ac.ae

Tip

Work in progress: The code will be updated as the research progresses.

Project Structure

filography/
├── analysis/          # Graph scripts for the benchmark report
├── config/            # Pipeline and recognition study settings
├── dataset/           # COCO val2017 images and labels (downloaded, not in git)
├── docs/              # Algorithm notes and findings
├── models/            # YOLO nano weights
├── native/            # threader C++ source and build script
├── recognition/       # Human recognition study: dataset builder, web app, scoring
├── reports/           # Benchmark reports and figures
├── src/               # Python benchmark pipeline
├── tests/             # Automated tests
├── build/             # Compiled threader (generated, not in git)
├── results/           # Experiment outputs (generated, not in git)
└── runs/              # YOLO training runs (generated, not in git)

Setup

Python 3.12 and Windows with the MSVC C++ build tools.

pip install -r requirements.txt
native\build.bat
python src/download_dataset.py -n 5000

native\build.bat compiles build\threader.exe.

Reproduce

Run the threading benchmark. This trains one YOLO model per algorithm and thread count, and needs a CUDA GPU:

python src/run_threaded_experiment.py

Add --daily-pause to stop GPU work between 15:00 and 23:00 each day. Results go to results/. Copy threaded_benchmark_results.md and .csv to reports/ to publish them, then redraw the figures:

python analysis/make_graphs.py
python analysis/plot_normalised.py

Threader output is deterministic: the same seed gives byte-identical PNGs.

Interactive terminal

Dataset, model selection and evaluation menus:

python src/run_terminal.py

Human recognition study

See recognition/README.md.

Tests

pytest tests

threading