update model with new weights

The new face detector is a EfficientDet implementation trained with
face images.

estimate_head_pose.py

@@ -1,16 +1,23 @@
 """Demo code shows how to estimate human head pose.
-Currently, human face is detected by a detector from an OpenCV DNN module.
-Then the face box is modified a little to suits the need of landmark
-detection. The facial landmark detection is done by a custom Convolutional
-Neural Network trained with TensorFlow. After that, head pose is estimated
-by solving a PnP problem.
+
+Three major steps for this code.
+
+Step 1: face detection. The human faces are detected by a deep learning face
+    face_detector .Then the face boxes are modified a little to suits the need of
+    landmark detection.
+Step 2: facial landmark detection. This is done by a custom Convolutional
+    Neural Network trained with TensorFlow.
+Step 3: head pose estimation. The pose is estimated by solving a PnP problem.
+
+All models and training code are available at: https://github.com/yinguobing/head-pose-estimation
 """
 from argparse import ArgumentParser
-from multiprocessing import Process, Queue
 
 import cv2
 import numpy as np
+import tensorflow as tf
 
+from face_detector import FaceDetector
 from mark_detector import MarkDetector
 from os_detector import detect_os
 from pose_estimator import PoseEstimator
@@ -18,12 +25,11 @@ from stabilizer import Stabilizer
 
 print("OpenCV version: {}".format(cv2.__version__))
 
-# multiprocessing may not work on Windows and macOS, check OS for safety.
-detect_os()
+devices = tf.config.list_physical_devices('GPU')
+for device in devices:
+    tf.config.experimental.set_memory_growth(device, True)
 
-CNN_INPUT_SIZE = 128
-
-# Take arguments from user input.
+# Parse arguments from user inputs.
 parser = ArgumentParser()
 parser.add_argument("--video", type=str, default=None,
                     help="Video file to be processed.")
@@ -32,41 +38,36 @@ parser.add_argument("--cam", type=int, default=None,
 args = parser.parse_args()
 
 
-def get_face(detector, img_queue, box_queue):
-    """Get face from image queue. This function is used for multiprocessing"""
-    while True:
-        image = img_queue.get()
-        box = detector.extract_cnn_facebox(image)
-        box_queue.put(box)
-
-
 def main():
-    """MAIN"""
-    # Video source from webcam or video file.
+    """Run human head pose estimation from video files."""
+
+    # What is the threshold value for face detection.
+    threshold = 0.5
+
+    # Setup the video source. If no video file provided, the default webcam will
+    # be used.
     video_src = args.cam if args.cam is not None else args.video
     if video_src is None:
         print("Warning: video source not assigned, default webcam will be used.")
         video_src = 0
 
     cap = cv2.VideoCapture(video_src)
+
+    # If reading frames from a webcam, try setting the camera resolution.
     if video_src == 0:
         cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
-    _, sample_frame = cap.read()
 
-    # Introduce mark_detector to detect landmarks.
-    mark_detector = MarkDetector()
+    # Get the real frame resolution.
+    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
 
-    # Setup process and queues for multiprocessing.
-    img_queue = Queue()
-    box_queue = Queue()
-    img_queue.put(sample_frame)
-    box_process = Process(target=get_face, args=(
-        mark_detector, img_queue, box_queue,))
-    box_process.start()
+    # Introduce a mark_face_detector to detect face marks.
+    detector_mark = MarkDetector("assets/mark_model")
 
-    # Introduce pose estimator to solve pose. Get one frame to setup the
-    # estimator according to the image size.
-    height, width = sample_frame.shape[:2]
+    # Introduce a face face_detector to detect human faces.
+    detector_face = FaceDetector("assets/face_model")
+
+    # Introduce pose estimator to solve pose.
     pose_estimator = PoseEstimator(img_size=(height, width))
 
     # Introduce scalar stabilizers for pose.
@@ -76,9 +77,13 @@ def main():
         cov_process=0.1,
         cov_measure=0.1) for _ in range(6)]
 
+    # Introduce a metter to measure the FPS.
     tm = cv2.TickMeter()
 
+    # Loop through the video frames.
     while True:
+        tm.start()
+
         # Read frame, crop it, flip it, suits your needs.
         frame_got, frame = cap.read()
         if frame_got is False:
@@ -91,38 +96,40 @@ def main():
         if video_src == 0:
             frame = cv2.flip(frame, 2)
 
-        # Pose estimation by 3 steps:
-        # 1. detect face;
-        # 2. detect landmarks;
-        # 3. estimate pose
+        # Preprocess the input image.
+        _image = detector_face.preprocess(frame)
 
-        # Feed frame to image queue.
-        img_queue.put(frame)
+        # Run the model
+        boxes, scores, classes = detector_face.predict(_image, threshold)
 
-        # Get face from box queue.
-        facebox = box_queue.get()
+        # Transform the boxes into squares.
+        boxes = detector_face.transform_to_square(
+            boxes, scale=1.22, offset=(0, 0.13))
 
-        if facebox is not None:
-            # Detect landmarks from image of 128x128.
-            face_img = frame[facebox[1]: facebox[3],
-                             facebox[0]: facebox[2]]
-            face_img = cv2.resize(face_img, (CNN_INPUT_SIZE, CNN_INPUT_SIZE))
-            face_img = cv2.cvtColor(face_img, cv2.COLOR_BGR2RGB)
+        # Clip the boxes if they cross the image boundaries.
+        boxes, _ = detector_face.clip_boxes(boxes, (0, 0, height, width))
 
-            tm.start()
-            marks = mark_detector.detect_marks(face_img)
-            tm.stop()
+        # Detect facial marks.
+        if boxes.size > 0:
+            # Get one face image.
+            facebox = boxes[0]
+            top, left, bottom, right = [int(x) for x in facebox]
+            face_image = frame[top:bottom, left:right]
+
+            # Run detection.
+            face_image = detector_mark.preprocess(face_image)
+            marks = detector_mark.predict(face_image)
 
             # Convert the marks locations from local CNN to global image.
-            marks *= (facebox[2] - facebox[0])
-            marks[:, 0] += facebox[0]
-            marks[:, 1] += facebox[1]
-
-            # Uncomment following line to show raw marks.
-            # mark_detector.draw_marks(frame, marks, color=(0, 255, 0))
+            marks *= (right - left)
+            marks[:, 0] += left
+            marks[:, 1] += top
 
             # Uncomment following line to show facebox.
-            # mark_detector.draw_box(frame, [facebox])
+            # detector_face.draw_box(frame, facebox, scores[0])
+
+            # Uncomment following line to show raw marks.
+            detector_mark.draw_marks(frame, marks, color=(0, 255, 0))
 
             # Try pose estimation with 68 points.
             pose = pose_estimator.solve_pose_by_68_points(marks)
@@ -144,17 +151,19 @@ def main():
             frame, steady_pose[0], steady_pose[1], color=(128, 255, 128))
 
             # Uncomment following line to draw head axes on frame.
-            # pose_estimator.draw_axes(frame, steady_pose[0], steady_pose[1])
+            pose_estimator.draw_axes(frame, steady_pose[0], steady_pose[1])
+
+        tm.stop()
+
+        # Draw FPS on the screen's top left corner.
+        cv2.putText(frame, "FPS: {:.0f}".format(tm.getFPS()), (24, 24),
+                    cv2.FONT_HERSHEY_DUPLEX, 0.8, (255, 255, 255), 1, cv2.LINE_AA)
 
         # Show preview.
         cv2.imshow("Preview", frame)
-        if cv2.waitKey(10) == 27:
+        if cv2.waitKey(1) == 27:
             break
 
-    # Clean up the multiprocessing process.
-    box_process.terminate()
-    box_process.join()
-
 
 if __name__ == '__main__':
     main()

face_detector.py

@@ -0,0 +1,161 @@
+"""Object detector based on EfficientDet model.
+
+This module supports inferenceing with the official EfficientDet model. 
+For more details: https://github.com/yinguobing/efficientdet-runner
+"""
+import cv2
+import numpy as np
+import tensorflow as tf
+
+
+class FaceDetector(object):
+    """Mini module to run the EfficientDet model."""
+
+    def __init__(self, saved_model):
+        """Build an EfficientDet model runner.
+
+        Args:
+            saved_model: the string path to the SavedModel.
+        """
+        self.scale_width = 0
+        self.scale_height = 0
+        self.input_size = 512
+
+        # Load the SavedModel object.
+        imported = tf.saved_model.load(saved_model)
+        self._predict_fn = imported.signatures["serving_default"]
+
+        # To avoid garbage collected by Python, see TensorFlow issue:37615
+        self._predict_fn._backref_to_saved_model = imported
+
+    def preprocess(self, image):
+        """Preprocess the input image."""
+
+        # Scale the image first.
+        height, width, _ = image.shape
+        self.ratio = self.input_size / max(height, width)
+        image = cv2.resize(
+            image, (int(self.ratio * width), int(self.ratio * height)))
+
+        # Convert to RGB.
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+
+        # Then pad the image to input size.
+        self.padding_h = self.input_size - int(self.ratio * width)
+        self.padding_v = self.input_size - int(self.ratio * height)
+        image = cv2.copyMakeBorder(
+            image, 0, self.padding_v, 0, self.padding_h, cv2.BORDER_CONSTANT, (0, 0, 0))
+
+        return image
+
+    def __filter(self, detections, threshold):
+        """Filter the detection results by score threshold."""
+        # Get the detection results.
+        boxes = detections['output_0'].numpy()[0]
+        scores = detections['output_1'].numpy()[0]
+        classes = detections['output_2'].numpy()[0]
+
+        # Filter out the results by score threshold.
+        mask = scores > threshold
+        boxes = boxes[mask]
+        scores = scores[mask]
+        classes = classes[mask]
+
+        return boxes, scores, classes
+
+    @tf.function
+    def _predict(self, images):
+        return self._predict_fn(images)
+
+    def predict(self, image, threshold):
+        """Run inference with image inputs.
+
+        Args:
+            image: a numpy array as an input image.
+
+        Returns:
+            predictions: result.
+        """
+        frame_tensor = tf.constant(image, dtype=tf.uint8)
+        frame_tensor = tf.expand_dims(frame_tensor, axis=0)
+        detections = self._predict(frame_tensor)
+        boxes, scores, classes = self.__filter(detections, threshold)
+
+        # Scale the box back to the original image size.
+        boxes /= self.ratio
+
+        # Crop out the padding area.
+        boxes[:, 2] = np.minimum(
+            boxes[:, 2], (self.input_size - self.padding_v)/self.ratio)
+        boxes[:, 1] = np.minimum(
+            boxes[:, 1], (self.input_size - self.padding_h)/self.ratio)
+
+        return boxes, scores, classes
+
+    def transform_to_square(self, boxes, scale=1.0, offset=(0, 0)):
+        """Get the square bounding boxes.
+
+        Args:
+            boxes: input boxes [[ymin, xmin, ymax, xmax], ...]
+            scale: ratio to scale the boxes
+            offset: a tuple of offset to move the boxes (x, y)
+
+        Returns:
+            square boxes.
+        """
+        ymins, xmins, ymaxs, xmaxs = np.split(boxes, 4, 1)
+        width = xmaxs - xmins
+        height = ymaxs - ymins
+
+        # How much to move.
+        offset_x = offset[0] * width
+        offset_y = offset[1] * height
+
+        # Where is the center location.
+        center_x = np.floor_divide(xmins + xmaxs, 2) + offset_x
+        center_y = np.floor_divide(ymins + ymaxs, 2) + offset_y
+
+        # Make them squares.
+        margin = np.floor_divide(np.maximum(height, width) * scale, 2)
+        boxes = np.concatenate((center_y-margin, center_x-margin,
+                                center_y+margin, center_x+margin), axis=1)
+
+        return boxes
+
+    def clip_boxes(self, boxes, margins):
+        """Clip the boxes to the safe margins.
+
+        Args:
+            boxes: input boxes [[ymin, xmin, ymax, xmax], ...].
+            margins: a tuple of 4 int (top, left, bottom, right) as safe margins.
+
+        Returns:
+            boxes: clipped boxes.
+            clip_mark: the mark of clipped sides.
+        """
+        top, left, bottom, right = margins
+
+        clip_mark = (boxes[:, 0] < top, boxes[:, 1] < left,
+                     boxes[:, 2] > bottom, boxes[:, 3] > right)
+
+        boxes[:, 0] = np.maximum(boxes[:, 0], top)
+        boxes[:, 1] = np.maximum(boxes[:, 1], left)
+        boxes[:, 2] = np.minimum(boxes[:, 2], bottom)
+        boxes[:, 3] = np.minimum(boxes[:, 3], right)
+
+        return boxes, clip_mark
+
+    def draw_box(self, image, box, score, color=(0, 255, 0)):
+        """Draw the bounding box.
+
+        Args:
+            boxes: the face box.
+            color: the color of the box.
+            scores: detection score.
+
+        """
+        y0, x0, y1, x1 = [int(b) for b in box]
+        cv2.rectangle(image, (x0, y0), (x1, y1), (0, 255, 0), 2)
+        cv2.putText(image, "Face:{:.2f}".format(score),
+                    (x0, y0-7), cv2.FONT_HERSHEY_DUPLEX, 0.5, color,
+                    1, cv2.LINE_AA)

mark_detector.py

@@ -5,163 +5,42 @@ import tensorflow as tf
 from tensorflow import keras
 
 
-class FaceDetector:
-    """Detect human face from image"""
-
-    def __init__(self,
-                 dnn_proto_text='assets/deploy.prototxt',
-                 dnn_model='assets/res10_300x300_ssd_iter_140000.caffemodel'):
-        """Initialization"""
-        self.face_net = cv2.dnn.readNetFromCaffe(dnn_proto_text, dnn_model)
-        self.detection_result = None
-
-    def get_faceboxes(self, image, threshold=0.5):
-        """
-        Get the bounding box of faces in image using dnn.
-        """
-        rows, cols, _ = image.shape
-
-        confidences = []
-        faceboxes = []
-
-        self.face_net.setInput(cv2.dnn.blobFromImage(
-            image, 1.0, (300, 300), (104.0, 177.0, 123.0), False, False))
-        detections = self.face_net.forward()
-
-        for result in detections[0, 0, :, :]:
-            confidence = result[2]
-            if confidence > threshold:
-                x_left_bottom = int(result[3] * cols)
-                y_left_bottom = int(result[4] * rows)
-                x_right_top = int(result[5] * cols)
-                y_right_top = int(result[6] * rows)
-                confidences.append(confidence)
-                faceboxes.append(
-                    [x_left_bottom, y_left_bottom, x_right_top, y_right_top])
-
-        self.detection_result = [faceboxes, confidences]
-
-        return confidences, faceboxes
-
-    def draw_all_result(self, image):
-        """Draw the detection result on image"""
-        for facebox, conf in self.detection_result:
-            cv2.rectangle(image, (facebox[0], facebox[1]),
-                          (facebox[2], facebox[3]), (0, 255, 0))
-            label = "face: %.4f" % conf
-            label_size, base_line = cv2.getTextSize(
-                label, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)
-
-            cv2.rectangle(image, (facebox[0], facebox[1] - label_size[1]),
-                          (facebox[0] + label_size[0],
-                           facebox[1] + base_line),
-                          (0, 255, 0), cv2.FILLED)
-            cv2.putText(image, label, (facebox[0], facebox[1]),
-                        cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0))
-
-
 class MarkDetector:
     """Facial landmark detector by Convolutional Neural Network"""
 
-    def __init__(self, saved_model='assets/pose_model'):
+    def __init__(self, saved_model):
         """Initialization"""
         # A face detector is required for mark detection.
-        self.face_detector = FaceDetector()
+        self.input_size = 128
 
-        self.cnn_input_size = 128
-        self.marks = None
+        # Load the SavedModel object.
+        imported = tf.saved_model.load(saved_model)
+        self._predict_fn = imported.signatures["serving_default"]
+        self._predict_fn._backref_to_saved_model = imported
 
-        # Restore model from the saved_model file.
-        self.model = keras.models.load_model(saved_model)
+    def preprocess(self, image):
+        """Preprocess the input images."""
+        image = cv2.resize(image, (self.input_size, self.input_size))
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
 
-    @staticmethod
-    def draw_box(image, boxes, box_color=(255, 255, 255)):
-        """Draw square boxes on image"""
-        for box in boxes:
-            cv2.rectangle(image,
-                          (box[0], box[1]),
-                          (box[2], box[3]), box_color, 3)
+        return image
 
-    @staticmethod
-    def move_box(box, offset):
-        """Move the box to direction specified by vector offset"""
-        left_x = box[0] + offset[0]
-        top_y = box[1] + offset[1]
-        right_x = box[2] + offset[0]
-        bottom_y = box[3] + offset[1]
-        return [left_x, top_y, right_x, bottom_y]
+    @tf.function
+    def _predict(self, images):
+        return self._predict_fn(image_input=images)
 
-    @staticmethod
-    def get_square_box(box):
-        """Get a square box out of the given box, by expanding it."""
-        left_x = box[0]
-        top_y = box[1]
-        right_x = box[2]
-        bottom_y = box[3]
-
-        box_width = right_x - left_x
-        box_height = bottom_y - top_y
-
-        # Check if box is already a square. If not, make it a square.
-        diff = box_height - box_width
-        delta = int(abs(diff) / 2)
-
-        if diff == 0:                   # Already a square.
-            return box
-        elif diff > 0:                  # Height > width, a slim box.
-            left_x -= delta
-            right_x += delta
-            if diff % 2 == 1:
-                right_x += 1
-        else:                           # Width > height, a short box.
-            top_y -= delta
-            bottom_y += delta
-            if diff % 2 == 1:
-                bottom_y += 1
-
-        # Make sure box is always square.
-        assert ((right_x - left_x) == (bottom_y - top_y)), 'Box is not square.'
-
-        return [left_x, top_y, right_x, bottom_y]
-
-    @staticmethod
-    def box_in_image(box, image):
-        """Check if the box is in image"""
-        rows = image.shape[0]
-        cols = image.shape[1]
-        return box[0] >= 0 and box[1] >= 0 and box[2] <= cols and box[3] <= rows
-
-    def extract_cnn_facebox(self, image):
-        """Extract face area from image."""
-        _, raw_boxes = self.face_detector.get_faceboxes(
-            image=image, threshold=0.9)
-
-        for box in raw_boxes:
-            # Move box down.
-            offset_y = int(abs((box[3] - box[1]) * 0.12))
-            box_moved = self.move_box(box, [0, offset_y])
-
-            # Make box square.
-            facebox = self.get_square_box(box_moved)
-
-            if self.box_in_image(facebox, image):
-                return facebox
-
-        return None
-
-    def detect_marks(self, images):
+    def predict(self, image):
         """Detect marks from images"""
 
-        # # Actual detection.
-        marks = self.model.predict(tf.expand_dims(images, axis=0))
+        # Actual detection.
+        marks = self._predict(tf.convert_to_tensor([image], dtype=tf.float32))
 
         # Convert predictions to landmarks.
-        marks = np.reshape(marks, (-1, 2))
+        marks = np.reshape(marks['dense_1'].numpy(), (-1, 2))
 
         return marks
 
-    @staticmethod
-    def draw_marks(image, marks, color=(255, 255, 255)):
+    def draw_marks(self, image, marks, color=(255, 255, 255)):
         """Draw mark points on image"""
         for mark in marks:
             cv2.circle(image, (int(mark[0]), int(