FOMO: Fast Object Localization

FOMO is a lightweight point localization model designed for edge AI applications. Instead of regressing bounding boxes, FOMO downsamples the input image (for example, mapping a 192x192 input to a 24x24 grid) and predicts class probabilities and coordinates on a per-cell basis.

Installation

Install the package via PyPI:

pip install fomo-edge-ai

Usage

# Model 
from fomo import FOMO
model = FOMO(model_path=None, size="s", nb_classes=1, device="cpu")

# Training

results = model.train(
    allow_experimental=True,
    data=str(data_yaml_path),   # YOLO style data.yaml
                                # Only bounding-box style datasets supported.
                                # TODO: add support for point-level annotation datasets
    epochs=EPOCHS,
    batch=BATCH,
    lr0=3e-4,
    eval_interval=1,
    workers=2,
    device=device,
    project=PROJECT,
    name=RUN_NAME,
    exist_ok=True,
    patience=0,
)

# Export as TFLite model

fp32_path = trained.export(output_path=str(weights_dir / f"{RUN_NAME}_fp32.tflite"))

# INT8 Quantization

int8_path = quantizer.quantize(
    fp32_tflite=fp32_path,
    calibration_data=calib_iter,
    config=config,
    output_path=str(weights_dir / f"{RUN_NAME}_int8.tflite")
)

Model Hosting

Models are currently available on Hugging Face:

https://huggingface.co/fomo-edge-ai/FOMO

Examples

Refer to examples/ for detailed examples on training and inference.

Tests

Tests are completed using modal. Install the modal cli and run

make test

License

Code is licensed under the Apache License 2.0. Pre-trained weights are hosted externally and may inherit separate licensing terms. Check details in the specific weight repositories.