add pose estimation model (#152)

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   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
+
+       http://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.

README.md

@@ -0,0 +1,34 @@
+# Pose estimation from MediaPipe Pose
+
+This model estimates 33 pose keypoints and person segmentation mask per detected person from [person detector](../person_detection_mediapipe). (The image below is referenced from [MediaPipe Pose Keypoints](https://github.com/tensorflow/tfjs-models/tree/master/pose-detection#blazepose-keypoints-used-in-mediapipe-blazepose))
+ 
+![MediaPipe Pose Landmark](examples/pose_landmarks.png)
+
+This model is converted from TFlite to ONNX using following tools:
+- TFLite model to ONNX: https://github.com/onnx/tensorflow-onnx
+- simplified by [onnx-simplifier](https://github.com/daquexian/onnx-simplifier)
+
+**Note**:
+- Visit https://github.com/google/mediapipe/blob/master/docs/solutions/models.md#pose for models of larger scale.
+## Demo
+
+Run the following commands to try the demo:
+```bash
+# detect on camera input
+python demo.py
+# detect on an image
+python demo.py -i /path/to/image -v
+```
+
+### Example outputs
+
+![webcam demo](examples/mpposeest_demo.webp)
+
+## License
+
+All files in this directory are licensed under [Apache 2.0 License](LICENSE).
+
+## Reference
+- MediaPipe Pose: https://developers.google.com/mediapipe/solutions/vision/pose_landmarker
+- MediaPipe pose model and model card: https://github.com/google/mediapipe/blob/master/docs/solutions/models.md#pose
+- BlazePose TFJS: https://github.com/tensorflow/tfjs-models/tree/master/pose-detection/src/blazepose_tfjs

demo.py

@@ -0,0 +1,252 @@
+import sys
+import argparse
+
+import numpy as np
+import cv2 as cv
+
+from mp_pose import MPPose
+
+sys.path.append('../person_detection_mediapipe')
+from mp_persondet import MPPersonDet
+
+# Check OpenCV version
+assert cv.__version__ >= "4.7.0", \
+       "Please install latest opencv-python to try this demo: python3 -m pip install --upgrade opencv-python"
+
+# Valid combinations of backends and targets
+backend_target_pairs = [
+    [cv.dnn.DNN_BACKEND_OPENCV, cv.dnn.DNN_TARGET_CPU],
+    [cv.dnn.DNN_BACKEND_CUDA,   cv.dnn.DNN_TARGET_CUDA],
+    [cv.dnn.DNN_BACKEND_CUDA,   cv.dnn.DNN_TARGET_CUDA_FP16],
+    [cv.dnn.DNN_BACKEND_TIMVX,  cv.dnn.DNN_TARGET_NPU],
+    [cv.dnn.DNN_BACKEND_CANN,   cv.dnn.DNN_TARGET_NPU]
+]
+
+parser = argparse.ArgumentParser(description='Pose Estimation from MediaPipe')
+parser.add_argument('--input', '-i', type=str,
+                    help='Path to the input image. Omit for using default camera.')
+parser.add_argument('--model', '-m', type=str, default='./pose_estimation_mediapipe_2023mar.onnx',
+                    help='Path to the model.')
+parser.add_argument('--backend_target', '-bt', type=int, default=0,
+                    help='''Choose one of the backend-target pair to run this demo:
+                        {:d}: (default) OpenCV implementation + CPU,
+                        {:d}: CUDA + GPU (CUDA),
+                        {:d}: CUDA + GPU (CUDA FP16),
+                        {:d}: TIM-VX + NPU,
+                        {:d}: CANN + NPU
+                    '''.format(*[x for x in range(len(backend_target_pairs))]))
+parser.add_argument('--conf_threshold', type=float, default=0.8,
+                    help='Filter out hands of confidence < conf_threshold.')
+parser.add_argument('--save', '-s', action='store_true',
+                    help='Specify to save results. This flag is invalid when using camera.')
+parser.add_argument('--vis', '-v', action='store_true',
+                    help='Specify to open a window for result visualization. This flag is invalid when using camera.')
+args = parser.parse_args()
+
+def visualize(image, poses):
+    display_screen = image.copy()
+    display_3d = np.zeros((400, 400, 3), np.uint8)
+    cv.line(display_3d, (200, 0), (200, 400), (255, 255, 255), 2)
+    cv.line(display_3d, (0, 200), (400, 200), (255, 255, 255), 2)
+    cv.putText(display_3d, 'Main View', (0, 12), cv.FONT_HERSHEY_DUPLEX, 0.5, (0, 0, 255))
+    cv.putText(display_3d, 'Top View', (200, 12), cv.FONT_HERSHEY_DUPLEX, 0.5, (0, 0, 255))
+    cv.putText(display_3d, 'Left View', (0, 212), cv.FONT_HERSHEY_DUPLEX, 0.5, (0, 0, 255))
+    cv.putText(display_3d, 'Right View', (200, 212), cv.FONT_HERSHEY_DUPLEX, 0.5, (0, 0, 255))
+    is_draw = False  # ensure only one person is drawn
+
+    def _draw_lines(image, landmarks, keep_landmarks, is_draw_point=True, thickness=2):
+
+        def _draw_by_presence(idx1, idx2):
+            if keep_landmarks[idx1] and keep_landmarks[idx2]:
+                cv.line(image, landmarks[idx1], landmarks[idx2], (255, 255, 255), thickness)
+
+        _draw_by_presence(0, 1)
+        _draw_by_presence(1, 2)
+        _draw_by_presence(2, 3)
+        _draw_by_presence(3, 7)
+        _draw_by_presence(0, 4)
+        _draw_by_presence(4, 5)
+        _draw_by_presence(5, 6)
+        _draw_by_presence(6, 8)
+
+        _draw_by_presence(9, 10)
+
+        _draw_by_presence(12, 14)
+        _draw_by_presence(14, 16)
+        _draw_by_presence(16, 22)
+        _draw_by_presence(16, 18)
+        _draw_by_presence(16, 20)
+        _draw_by_presence(18, 20)
+
+        _draw_by_presence(11, 13)
+        _draw_by_presence(13, 15)
+        _draw_by_presence(15, 21)
+        _draw_by_presence(15, 19)
+        _draw_by_presence(15, 17)
+        _draw_by_presence(17, 19)
+
+        _draw_by_presence(11, 12)
+        _draw_by_presence(11, 23)
+        _draw_by_presence(23, 24)
+        _draw_by_presence(24, 12)
+
+        _draw_by_presence(24, 26)
+        _draw_by_presence(26, 28)
+        _draw_by_presence(28, 30)
+        _draw_by_presence(28, 32)
+        _draw_by_presence(30, 32)
+
+        _draw_by_presence(23, 25)
+        _draw_by_presence(25, 27)
+        _draw_by_presence(27, 31)
+        _draw_by_presence(27, 29)
+        _draw_by_presence(29, 31)
+
+        if is_draw_point:
+            for i, p in enumerate(landmarks):
+                if keep_landmarks[i]:
+                    cv.circle(image, p, thickness, (0, 0, 255), -1)
+
+    for idx, pose in enumerate(poses):
+        bbox, landmarks_screen, landmarks_word, mask, heatmap, conf = pose
+
+        edges = cv.Canny(mask, 100, 200)
+        kernel = np.ones((2, 2), np.uint8) # expansion edge to 2 pixels
+        edges = cv.dilate(edges, kernel, iterations=1)
+        edges_bgr = cv.cvtColor(edges, cv.COLOR_GRAY2BGR)
+        edges_bgr[edges == 255] = [0, 255, 0]
+        display_screen = cv.add(edges_bgr, display_screen)
+
+
+        # draw box
+        bbox = bbox.astype(np.int32)
+        cv.rectangle(display_screen, bbox[0], bbox[1], (0, 255, 0), 2)
+        cv.putText(display_screen, '{:.4f}'.format(conf), (bbox[0][0], bbox[0][1] + 12), cv.FONT_HERSHEY_DUPLEX, 0.5, (0, 0, 255))
+        # Draw line between each key points
+        landmarks_screen = landmarks_screen[:-6, :]
+        landmarks_word = landmarks_word[:-6, :]
+
+        keep_landmarks = landmarks_screen[:, 4] > 0.8 # only show visible keypoints which presence bigger than 0.8
+
+        landmarks_screen = landmarks_screen
+        landmarks_word = landmarks_word
+
+        landmarks_xy = landmarks_screen[:, 0: 2].astype(np.int32)
+        _draw_lines(display_screen, landmarks_xy, keep_landmarks, is_draw_point=False)
+
+        # z value is relative to HIP, but we use constant to instead
+        for i, p in enumerate(landmarks_screen[:, 0: 3].astype(np.int32)):
+            if keep_landmarks[i]:
+                cv.circle(display_screen, np.array([p[0], p[1]]), 2, (0, 0, 255), -1)
+
+        if is_draw is False:
+            is_draw = True
+            # Main view
+            landmarks_xy = landmarks_word[:, [0, 1]]
+            landmarks_xy = (landmarks_xy * 100 + 100).astype(np.int32)
+            _draw_lines(display_3d, landmarks_xy, keep_landmarks, thickness=2)
+
+            # Top view
+            landmarks_xz = landmarks_word[:, [0, 2]]
+            landmarks_xz[:, 1] = -landmarks_xz[:, 1]
+            landmarks_xz = (landmarks_xz * 100 + np.array([300, 100])).astype(np.int32)
+            _draw_lines(display_3d, landmarks_xz,keep_landmarks, thickness=2)
+
+            # Left view
+            landmarks_yz = landmarks_word[:, [2, 1]]
+            landmarks_yz[:, 0] = -landmarks_yz[:, 0]
+            landmarks_yz = (landmarks_yz * 100 + np.array([100, 300])).astype(np.int32)
+            _draw_lines(display_3d, landmarks_yz, keep_landmarks, thickness=2)
+
+            # Right view
+            landmarks_zy = landmarks_word[:, [2, 1]]
+            landmarks_zy = (landmarks_zy * 100 + np.array([300, 300])).astype(np.int32)
+            _draw_lines(display_3d, landmarks_zy, keep_landmarks, thickness=2)
+
+    return display_screen, display_3d
+
+if __name__ == '__main__':
+    backend_id = backend_target_pairs[args.backend_target][0]
+    target_id = backend_target_pairs[args.backend_target][1]
+
+    # person detector
+    person_detector = MPPersonDet(modelPath='../person_detection_mediapipe/person_detection_mediapipe_2023mar.onnx',
+                                  nmsThreshold=0.3,
+                                  scoreThreshold=0.5,
+                                  topK=5000, # usually only one person has good performance
+                                  backendId=backend_id,
+                                  targetId=target_id)
+    # pose estimator
+    pose_estimator = MPPose(modelPath=args.model,
+                            confThreshold=args.conf_threshold,
+                            backendId=backend_id,
+                            targetId=target_id)
+
+    # If input is an image
+    if args.input is not None:
+        image = cv.imread(args.input)
+
+        # person detector inference
+        persons = person_detector.infer(image)
+        poses = []
+
+        # Estimate the pose of each person
+        for person in persons:
+            # pose estimator inference
+            pose = pose_estimator.infer(image, person)
+            if pose is not None:
+                poses.append(pose)
+        # Draw results on the input image
+        image, view_3d = visualize(image, poses)
+
+        if len(persons) == 0:
+            print('No person detected!')
+        else:
+            print('Person detected!')
+
+        # Save results
+        if args.save:
+            cv.imwrite('result.jpg', image)
+            print('Results saved to result.jpg\n')
+
+        # Visualize results in a new window
+        if args.vis:
+            cv.namedWindow(args.input, cv.WINDOW_AUTOSIZE)
+            cv.imshow(args.input, image)
+            cv.imshow('3D Pose Demo', view_3d)
+            cv.waitKey(0)
+    else:  # Omit input to call default camera
+        deviceId = 0
+        cap = cv.VideoCapture(deviceId)
+
+        tm = cv.TickMeter()
+        while cv.waitKey(1) < 0:
+            hasFrame, frame = cap.read()
+            if not hasFrame:
+                print('No frames grabbed!')
+                break
+
+            # person detector inference
+            persons = person_detector.infer(frame)
+            poses = []
+
+            tm.start()
+            # Estimate the pose of each person
+            for person in persons:
+                # pose detector inference
+                pose = pose_estimator.infer(frame, person)
+                if pose is not None:
+                    poses.append(pose)
+            tm.stop()
+            # Draw results on the input image
+            frame, view_3d = visualize(frame, poses)
+
+            if len(persons) == 0:
+                print('No person detected!')
+            else:
+                print('Person detected!')
+                cv.putText(frame, 'FPS: {:.2f}'.format(tm.getFPS()), (0, 15), cv.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255))
+
+            cv.imshow('MediaPipe Pose Detection Demo', frame)
+            cv.imshow('3D Pose Demo', view_3d)
+            tm.reset()

mp_pose.py

@@ -0,0 +1,179 @@
+import numpy as np
+import cv2 as cv
+
+class MPPose:
+    def __init__(self, modelPath, confThreshold=0.5, backendId=0, targetId=0):
+        self.model_path = modelPath
+        self.conf_threshold = confThreshold
+        self.backend_id = backendId
+        self.target_id = targetId
+
+        self.input_size = np.array([256, 256])  # wh
+        # RoI will be larger so the performance will be better, but preprocess will be slower. Default to 1.
+        self.PERSON_BOX_PRE_ENLARGE_FACTOR = 1
+        self.PERSON_BOX_ENLARGE_FACTOR = 1.25
+
+        self.model = cv.dnn.readNet(self.model_path)
+        self.model.setPreferableBackend(self.backend_id)
+        self.model.setPreferableTarget(self.target_id)
+
+    @property
+    def name(self):
+        return self.__class__.__name__
+
+    def setBackendAndTarget(self, backendId, targetId):
+        self._backendId = backendId
+        self._targetId = targetId
+        self.model.setPreferableBackend(self.backend_id)
+        self.model.setPreferableTarget(self.target_id)
+
+    def _preprocess(self, image, person):
+        '''
+        Rotate input for inference.
+        Parameters:
+          image - input image of BGR channel order
+          face_bbox - human face bounding box found in image of format [[x1, y1], [x2, y2]] (top-left and bottom-right points)
+          person_landmarks - 4 landmarks (2 full body points, 2 upper body points) of shape [4, 2]
+        Returns:
+          rotated_person - rotated person image for inference
+          rotate_person_bbox - person box of interest range
+          angle - rotate angle for person
+          rotation_matrix - matrix for rotation and de-rotation
+          pad_bias - pad pixels of interest range
+        '''
+        # crop and pad image to interest range
+        pad_bias = np.array([0, 0], dtype=np.int32)  # left, top
+        person_keypoints = person[4: 12].reshape(-1, 2)
+        mid_hip_point = person_keypoints[0]
+        full_body_point = person_keypoints[1]
+        # get RoI
+        full_dist = np.linalg.norm(mid_hip_point - full_body_point)
+        full_bbox = np.array([mid_hip_point - full_dist, mid_hip_point + full_dist], np.int32)
+        # enlarge to make sure full body can be cover
+        center_bbox = np.sum(full_bbox, axis=0) / 2
+        wh_bbox = full_bbox[1] - full_bbox[0]
+        new_half_size = wh_bbox * self.PERSON_BOX_PRE_ENLARGE_FACTOR / 2
+        full_bbox = np.array([
+            center_bbox - new_half_size,
+            center_bbox + new_half_size], np.int32)
+
+        person_bbox = full_bbox.copy()
+        # refine person bbox
+        person_bbox[:, 0] = np.clip(person_bbox[:, 0], 0, image.shape[1])
+        person_bbox[:, 1] = np.clip(person_bbox[:, 1], 0, image.shape[0])
+        # crop to the size of interest
+        image = image[person_bbox[0][1]:person_bbox[1][1], person_bbox[0][0]:person_bbox[1][0], :]
+        # pad to square
+        left, top = person_bbox[0] - full_bbox[0]
+        right, bottom = full_bbox[1] - person_bbox[1]
+        image = cv.copyMakeBorder(image, top, bottom, left, right, cv.BORDER_CONSTANT, None, (0, 0, 0))
+        pad_bias += person_bbox[0] - [left, top]
+        # compute rotation
+        mid_hip_point -= pad_bias
+        full_body_point -= pad_bias
+        radians = np.pi / 2 - np.arctan2(-(full_body_point[1] - mid_hip_point[1]), full_body_point[0] - mid_hip_point[0])
+        radians = radians - 2 * np.pi * np.floor((radians + np.pi) / (2 * np.pi))
+        angle = np.rad2deg(radians)
+        #  get rotation matrix
+        rotation_matrix = cv.getRotationMatrix2D(mid_hip_point, angle, 1.0)
+        #  get rotated image
+        rotated_image = cv.warpAffine(image, rotation_matrix, (image.shape[1], image.shape[0]))
+        #  get landmark bounding box
+        blob = cv.resize(rotated_image, dsize=self.input_size, interpolation=cv.INTER_AREA).astype(np.float32)
+        rotated_person_bbox = np.array([[0, 0], [image.shape[1], image.shape[0]]], dtype=np.int32)
+        blob = cv.cvtColor(blob, cv.COLOR_BGR2RGB)
+        blob = blob / 255. # [0, 1]
+        return blob[np.newaxis, :, :, :], rotated_person_bbox, angle, rotation_matrix, pad_bias
+
+    def infer(self, image, person):
+        h, w, _ = image.shape
+        # Preprocess
+        input_blob, rotated_person_bbox, angle, rotation_matrix, pad_bias = self._preprocess(image, person)
+
+        # Forward
+        self.model.setInput(input_blob)
+        output_blob = self.model.forward(self.model.getUnconnectedOutLayersNames())
+
+        # Postprocess
+        results = self._postprocess(output_blob, rotated_person_bbox, angle, rotation_matrix, pad_bias, np.array([w, h]))
+        return results # [bbox_coords, landmarks_coords, conf]
+
+    def _postprocess(self, blob, rotated_person_bbox, angle, rotation_matrix, pad_bias, img_size):
+        landmarks, conf, mask, heatmap, landmarks_word = blob
+
+        conf = conf[0][0]
+        if conf < self.conf_threshold:
+            return None
+
+        landmarks = landmarks[0].reshape(-1, 5)  # shape: (1, 195) -> (39, 5)
+        landmarks_word = landmarks_word[0].reshape(-1, 3) # shape: (1, 117) -> (39, 3)
+
+        # recover sigmoid score
+        landmarks[:, 3:] = 1 / (1 + np.exp(-landmarks[:, 3:]))
+        # TODO: refine landmarks with heatmap. reference: https://github.com/tensorflow/tfjs-models/blob/master/pose-detection/src/blazepose_tfjs/detector.ts#L577-L582
+        heatmap = heatmap[0]
+
+        # transform coords back to the input coords
+        wh_rotated_person_bbox = rotated_person_bbox[1] - rotated_person_bbox[0]
+        scale_factor = wh_rotated_person_bbox / self.input_size
+        landmarks[:, :2] = (landmarks[:, :2] - self.input_size / 2) * scale_factor
+        landmarks[:, 2] = landmarks[:, 2] * max(scale_factor) # depth scaling
+        coords_rotation_matrix = cv.getRotationMatrix2D((0, 0), angle, 1.0)
+        rotated_landmarks = np.dot(landmarks[:, :2], coords_rotation_matrix[:, :2])
+        rotated_landmarks = np.c_[rotated_landmarks, landmarks[:, 2:]]
+        rotated_landmarks_world = np.dot(landmarks_word[:, :2], coords_rotation_matrix[:, :2])
+        rotated_landmarks_world = np.c_[rotated_landmarks_world, landmarks_word[:, 2]]
+        # invert rotation
+        rotation_component = np.array([
+            [rotation_matrix[0][0], rotation_matrix[1][0]],
+            [rotation_matrix[0][1], rotation_matrix[1][1]]])
+        translation_component = np.array([
+            rotation_matrix[0][2], rotation_matrix[1][2]])
+        inverted_translation = np.array([
+            -np.dot(rotation_component[0], translation_component),
+            -np.dot(rotation_component[1], translation_component)])
+        inverse_rotation_matrix = np.c_[rotation_component, inverted_translation]
+        #  get box center
+        center = np.append(np.sum(rotated_person_bbox, axis=0) / 2, 1)
+        original_center = np.array([
+            np.dot(center, inverse_rotation_matrix[0]),
+            np.dot(center, inverse_rotation_matrix[1])])
+        landmarks[:, :2] = rotated_landmarks[:, :2] + original_center + pad_bias
+
+        # get bounding box from rotated_landmarks
+        bbox = np.array([
+            np.amin(landmarks[:, :2], axis=0),
+            np.amax(landmarks[:, :2], axis=0)])  # [top-left, bottom-right]
+        center_bbox = np.sum(bbox, axis=0) / 2
+        wh_bbox = bbox[1] - bbox[0]
+        new_half_size = wh_bbox * self.PERSON_BOX_ENLARGE_FACTOR / 2
+        bbox = np.array([
+            center_bbox - new_half_size,
+            center_bbox + new_half_size])
+
+        # invert rotation for mask
+        mask = mask[0].reshape(256, 256) # shape: (1, 256, 256, 1) -> (256, 256)
+        invert_rotation_matrix = cv.getRotationMatrix2D((mask.shape[1]/2, mask.shape[0]/2), -angle, 1.0)
+        invert_rotation_mask = cv.warpAffine(mask, invert_rotation_matrix, (mask.shape[1], mask.shape[0]))
+        # enlarge mask
+        invert_rotation_mask = cv.resize(invert_rotation_mask, wh_rotated_person_bbox)
+        # crop and pad mask
+        min_w, min_h = -np.minimum(pad_bias, 0)
+        left, top = np.maximum(pad_bias, 0)
+        pad_over = img_size - [invert_rotation_mask.shape[1], invert_rotation_mask.shape[0]] - pad_bias
+        max_w, max_h = np.minimum(pad_over, 0) + [invert_rotation_mask.shape[1], invert_rotation_mask.shape[0]]
+        right, bottom = np.maximum(pad_over, 0)
+        invert_rotation_mask = invert_rotation_mask[min_h:max_h, min_w:max_w]
+        invert_rotation_mask = cv.copyMakeBorder(invert_rotation_mask, top, bottom, left, right, cv.BORDER_CONSTANT, None, 0)
+        # binarize mask
+        invert_rotation_mask = np.where(invert_rotation_mask > 0, 255, 0).astype(np.uint8)
+
+        # 2*2 person bbox: [[x1, y1], [x2, y2]]
+        # 39*5 screen landmarks: 33 keypoints and 6 auxiliary points with [x, y, z, visibility, presence], z value is relative to HIP
+        # Visibility is probability that a keypoint is located within the frame and not occluded by another bigger body part or another object
+        # Presence is probability that a keypoint is located within the frame
+        # 39*3 world landmarks: 33 keypoints and 6 auxiliary points with [x, y, z] 3D metric x, y, z coordinate
+        # img_height*img_width mask: gray mask, where 255 indicates the full body of a person and 0 means background
+        # 64*64*39 heatmap: currently only used for refining landmarks, requires sigmod processing before use
+        # conf: confidence of prediction
+        return [bbox, landmarks, rotated_landmarks_world, invert_rotation_mask, heatmap, conf]