Google coral support (rebased onto master)

.github/workflows/unit-tests-on-python.yml

@@ -25,7 +25,7 @@ jobs:
         sudo apt-get update && sudo apt-get install -y libjpeg-dev libpng-dev \
           libtiff-dev libavformat-dev libpq-dev libfreeimage3
         python -m pip install --upgrade pip
-        python -m pip --no-cache-dir install -r requirements.txt
+        python -m pip --no-cache-dir install -r requirements.txt -e srcext/mtcnn_tflite/
         python -m src.services.facescan.plugins.setup
     - name: Test with pytest
       working-directory: ./embedding-calculator/

embedding-calculator/Dockerfile

@@ -61,7 +61,7 @@ ARG GPU_IDX=-1
 ENV GPU_IDX=$GPU_IDX INTEL_OPTIMIZATION=$INTEL_OPTIMIZATION
 ARG FACE_DETECTION_PLUGIN="facenet.FaceDetector"
 ARG CALCULATION_PLUGIN="facenet.Calculator"
-ARG EXTRA_PLUGINS="facenet.LandmarksDetector,agegender.AgeDetector,agegender.GenderDetector,facenet.facemask.MaskDetector,facenet.PoseEstimator"
+ARG EXTRA_PLUGINS="facenet.LandmarksDetector,agegender.AgeDetector,agegender.GenderDetector,facenet.facemask.MaskDetector"
 ENV FACE_DETECTION_PLUGIN=$FACE_DETECTION_PLUGIN CALCULATION_PLUGIN=$CALCULATION_PLUGIN \
     EXTRA_PLUGINS=$EXTRA_PLUGINS
 COPY src src

embedding-calculator/src/constants.py

@@ -25,8 +25,7 @@ class ENV(Constants):
 
     FACE_DETECTION_PLUGIN = get_env('FACE_DETECTION_PLUGIN', 'facenet.FaceDetector')
     CALCULATION_PLUGIN = get_env('CALCULATION_PLUGIN', 'facenet.Calculator')
-    EXTRA_PLUGINS = get_env_split('EXTRA_PLUGINS', 'facenet.LandmarksDetector,agegender.AgeDetector,agegender.GenderDetector,facenet.facemask.MaskDetector,facenet.PoseEstimator')
-
+    EXTRA_PLUGINS = get_env_split('EXTRA_PLUGINS', 'facenet.LandmarksDetector,agegender.AgeDetector,agegender.GenderDetector,facenet.facemask.MaskDetector,facenet.PoseEstimator,facenet.coralmtcnn.FaceDetector,facenet.coralmtcnn.Calculator')
     LOGGING_LEVEL_NAME = get_env('LOGGING_LEVEL_NAME', 'debug').upper()
     IS_DEV_ENV = get_env('FLASK_ENV', 'production') == 'development'
     BUILD_VERSION = get_env('APP_VERSION_STRING', 'dev')

embedding-calculator/src/services/facescan/plugins/facenet/coralmtcnn/__init__.py

@@ -0,0 +1,15 @@
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#       https://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
+#  or implied. See the License for the specific language governing
+#  permissions and limitations under the License.
+
+from src.services.facescan.plugins.dependencies import get_tensorflow
+
+requirements = get_tensorflow()

embedding-calculator/src/services/facescan/plugins/facenet/coralmtcnn/coralmtcnn.py

@@ -0,0 +1,170 @@
+#  Copyright (c) 2020 the original author or authors
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#       https://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
+#  or implied. See the License for the specific language governing
+#  permissions and limitations under the License.
+
+import logging
+import math
+import cv2
+from typing import List
+
+import tensorflow as tf
+import numpy as np
+from cached_property import cached_property
+from mtcnn_tflite.MTCNN import MTCNN
+
+from src.constants import ENV
+from src.services.dto.bounding_box import BoundingBoxDTO
+from src.services.facescan.plugins import mixins
+from src.services.facescan.imgscaler.imgscaler import ImgScaler
+from src.services.imgtools.proc_img import crop_img, squish_img
+from src.services.imgtools.types import Array3D
+from src.services.utils.pyutils import get_current_dir
+
+from src.services.facescan.plugins import base
+
+CURRENT_DIR = get_current_dir(__file__)
+logger = logging.getLogger(__name__)
+
+def prewhiten(img):
+    """ Normalize image."""
+    mean = np.mean(img)
+    std = np.std(img)
+    std_adj = np.maximum(std, 1.0 / np.sqrt(img.size))
+    y = np.multiply(np.subtract(img, mean), 1 / std_adj)
+    return y
+
+class FaceDetector(mixins.FaceDetectorMixin, base.BasePlugin):
+    FACE_MIN_SIZE = 20
+    SCALE_FACTOR = 0.709
+    BOX_MARGIN = 32
+    IMAGE_SIZE = 160
+    IMG_LENGTH_LIMIT = ENV.IMG_LENGTH_LIMIT
+
+    # detection settings
+    det_prob_threshold = 0.85
+    det_threshold_a = 0.9436513301
+    det_threshold_b = 0.7059968943
+    det_threshold_c = 0.5506904359
+
+    @cached_property
+    def _face_detection_net(self):
+        return MTCNN(
+            min_face_size=self.FACE_MIN_SIZE,
+            scale_factor=self.SCALE_FACTOR,
+            steps_threshold=[self.det_threshold_a, self.det_threshold_b, self.det_threshold_c]
+        )
+
+    def crop_face(self, img: Array3D, box: BoundingBoxDTO) -> Array3D:
+        return cv2.resize(crop_img(img, box), (self.IMAGE_SIZE, self.IMAGE_SIZE))
+
+    def find_faces(self, img: Array3D, det_prob_threshold: float = None) -> List[BoundingBoxDTO]:
+        if det_prob_threshold is None:
+            det_prob_threshold = self.det_prob_threshold
+        assert 0 <= det_prob_threshold <= 1
+        scaler = ImgScaler(self.IMG_LENGTH_LIMIT)
+        img = scaler.downscale_img(img)
+
+        fdn = self._face_detection_net
+        detect_face_result = fdn.detect_faces(img)
+        img_size = np.asarray(img.shape)[0:2]
+        bounding_boxes = []
+
+        for face in detect_face_result:
+            x, y, w, h = face['box']
+            margin_x = w / 8
+            margin_y = h / 8
+            box = BoundingBoxDTO(
+                x_min=int(np.maximum(x - margin_x, 0)),
+                y_min=int(np.maximum(y - margin_y, 0)),
+                x_max=int(np.minimum(x + w + margin_x, img_size[1])),
+                y_max=int(np.minimum(y + h + margin_y, img_size[0])),
+                np_landmarks=np.array([list(value) for value in face['keypoints'].values()]),
+                probability=face['confidence']
+            )
+            logger.debug(f"Found: {box}")
+            bounding_boxes.append(box)
+
+        filtered_bounding_boxes = []
+        for box in bounding_boxes:
+            box = box.scaled(scaler.upscale_coefficient)
+            if box.probability <= det_prob_threshold:
+                logger.debug(f'Box filtered out because below threshold ({det_prob_threshold}): {box}')
+                continue
+            filtered_bounding_boxes.append(box)
+        return filtered_bounding_boxes
+
+
+class Calculator(mixins.CalculatorMixin, base.BasePlugin):
+    ml_models = (
+        # converted facenet .tflite model
+        ('20180402-114759-edgetpu', '1Uwv8w6Uj5M_xdJI9sjay_wkoFoI_zbjk', (1.1817961, 5.291995557), 0.4),
+    )
+    BATCH_SIZE = 25
+    DELIGATES = 'libedgetpu.so.1'
+
+    @property
+    def ml_model_file(self):
+        return str(self.ml_model.path / f'{self.ml_model.name}.tflite')
+
+    @cached_property
+    def _embedding_calculator_tpu(self):
+        delegate_list = tf.lite.experimental.load_delegate(self.DELIGATES)
+        model = tf.lite.Interpreter(
+                                model_path=self.ml_model_file,
+                                experimental_delegates=[delegate_list])
+        return model
+
+    @cached_property
+    def _embedding_calculator(self):
+        model = tf.lite.Interpreter(model_path=self.ml_model_file)
+        return model
+
+    def calc_embedding(self, face_img: Array3D, mode='CPU') -> Array3D:
+        return self._calculate_embeddings([face_img], mode)[0]
+
+    def _calculate_embeddings(self, cropped_images, mode='CPU'):
+        """Run forward pass to calculate embeddings"""
+        if mode == 'TPU':
+            calc_model = self._embedding_calculator_tpu
+        else:
+            calc_model = self._embedding_calculator
+            cropped_images = [prewhiten(img).astype(np.float32) for img in cropped_images]
+
+        input_details = calc_model.get_input_details()
+        input_index = input_details[0]['index']
+        input_shape = input_details[0]['shape']
+        input_size = tuple(input_shape[1:4])
+
+        output_details = calc_model.get_output_details()
+        output_index = output_details[0]['index']
+        embedding_size = output_details[0]['shape'][1]
+
+        image_count = len(cropped_images)
+        batches_per_epoch = int(math.ceil(1.0 * image_count / self.BATCH_SIZE))
+        embeddings = np.zeros((image_count, embedding_size))
+        preprocessed_images = np.array([img for img in cropped_images])
+
+        for i in range(batches_per_epoch):
+            start_index = i * self.BATCH_SIZE
+            end_index = min((i + 1) * self.BATCH_SIZE, image_count)
+            calc_model.resize_tensor_input(input_index, (end_index-start_index, input_size[0], input_size[1], input_size[2]))
+            calc_model.resize_tensor_input(output_index, (end_index-start_index, embedding_size))
+            calc_model.allocate_tensors()
+            calc_model.set_tensor(input_index, preprocessed_images[start_index:end_index])
+            calc_model.invoke()
+            embeddings[start_index:end_index, :] = calc_model.get_tensor(output_index)
+        return embeddings
+
+
+class LandmarksDetector(mixins.LandmarksDetectorMixin, base.BasePlugin):
+    """ Extract landmarks from FaceDetector results."""

embedding-calculator/srcext/mtcnn_tflite/LICENSE

@@ -0,0 +1,22 @@
+MIT License
+
+Copyright (c) 2019 Iván de Paz Centeno
+Copyright (c) 2021 CDL Digidow <https://www.digidow.eu>
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

embedding-calculator/srcext/mtcnn_tflite/MANIFEST.in

@@ -0,0 +1 @@
+include mtcnn_tflite/data/mtcnn_weights.npy

embedding-calculator/srcext/mtcnn_tflite/README.md

@@ -0,0 +1,51 @@
+# MTCNN face recognition
+
+Implementation of the [MTCNN face detection algorithm](https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=7553523). This project converted the code from [ipazc/mtcnn](https://github.com/ipazc/mtcnn) to TF Lite.
+
+## Installation
+
+You can install the package through pip:
+
+```
+pip install mtcnn-tflite
+```
+
+## Quick start
+
+Similar to [the original implementation](https://github.com/ipazc/mtcnn), the following example illustrates the ease of use of this package:
+
+```
+>>> from mtcnn_tflite.MTCNN import MTCNN
+>>> import cv2
+>>>
+>>> img = cv2.cvtColor(cv2.imread("ivan.jpg"), cv2.COLOR_BGR2RGB)
+>>> detector = MTCNN()
+>>> detector.detect_faces(img)
+[
+    {
+        'box': [276, 88, 51, 68],
+        'confidence': 0.9989245533943176,
+        'keypoints': {
+            'left_eye': (291, 117),
+            'right_eye': (314, 114),
+            'nose': (303, 130),
+            'mouth_left': (296, 143),
+            'mouth_right': (314, 141)
+        }
+    }
+]
+```
+
+
+## Benchmark
+
+| Image size | TF version                            | Process time * |
+|------------|---------------------------------------|----------------|
+| 561x561    | [TF2](https://github.com/ipazc/mtcnn) | 698ms          |
+| 561x561    | **This repository** (TF Lite)         | 445ms          |
+
+\* executed on a CPU: Intel i7-10510U
+
+## License
+
+[MIT License](https://github.com/mobilesec/mtcnn-tflite/blob/master/LICENSE)