eDifFIQA ONNX

ONNX exports of the eDifFIQA face image quality models.

These models estimate the visual quality of a face image. They can be used before face recognition or face verification to filter low-quality face crops.

Note: The models expect cropped/aligned face images.

Export Details

Models were exported with torch==2.10.0 and simplified with onnxslim==0.1.93. All models were exported with dynamic batch size support and ONNX opset 20.

Available Models

Variant	File	Size
eDifFIQA-T	ediffiqa_t.onnx	tiny
eDifFIQA-S	ediffiqa_s.onnx	small
eDifFIQA-M	ediffiqa_m.onnx	medium
eDifFIQA-L	ediffiqa_l.onnx	large

Usage

from pathlib import Path

import cv2
import numpy as np
import onnxruntime as ort


class EDifFIQAOnnx:

    def __init__(
        self,
        model_path: Path | str,
        input_size: tuple[int, int] = (112, 112),
        providers: list[str] | None = None,
    ) -> None:
        self.model_path = str(model_path)
        self.input_size = input_size

        self.session = ort.InferenceSession(
            self.model_path,
            providers=providers or ["CPUExecutionProvider"],
        )

        self.input_name = self.session.get_inputs()[0].name
        self.output_name = self.session.get_outputs()[0].name

    def infer(self, image_bgr: np.ndarray) -> float:
        """
        Run quality inference on a BGR-aligned face image.

        :param image_bgr: BGR face image.
        :return: Quality score.
        """
        tensor = self._preprocess(image_bgr)
        output = self.session.run([self.output_name], {self.input_name: tensor})[0]
        return float(np.squeeze(output))

    def _preprocess(self, image_bgr: np.ndarray) -> np.ndarray:
        """
        Preprocess image.

        :param image_bgr: BGR image.
        :return: Input tensor with shape (1, 3, 112, 112).
        """
        image_rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
        image_rgb = cv2.resize(image_rgb, self.input_size)

        image = image_rgb.astype(np.float32)
        image = ((image / 255.0) - 0.5) / 0.5
        image = np.transpose(image, (2, 0, 1))[None, ...]

        return image.astype(np.float32)