lib_openpose_detector.py

'''
Openpose detector class.
'''

import sys
import cv2
import os
from sys import platform
import argparse
import time
import numpy as np
from tempfile import TemporaryFile

if True:  # Import openpose
    OPENPOSE_PYTHONPATH = os.environ['OPENPOSE_PYTHONPATH']
    sys.path.append(OPENPOSE_PYTHONPATH)
    from openpose import pyopenpose as op
    ROOT = os.path.dirname(os.path.abspath(__file__))+'/'

''' -------------------------------------- Settings -------------------------------------- '''

OPENPOSE_HOME = os.environ['OPENPOSE_HOME'] + "/"
MODEL_PATH = OPENPOSE_HOME + "models/"

''' ------------------------------- Command line arguments ------------------------------- '''


class OpenposeDetector(object):
    def __init__(self, keys_params={}):
        self._params = self.set_default_params()
        self.set_params(keys_params=keys_params)
        assert(self._params["face"] == False)  # I haven't added this.
        self._opWrapper = op.WrapperPython()
        self._opWrapper.configure(self._params)
        self._opWrapper.start()  # Start Openpose.
        self._datum = None

    def detect(self, color_image, is_return_joints=False):
        '''
        Arguments:
            color_image {image}: shape=[rows, cols, 3]; type=np.uint8.
        Return:
            datum {op.Datum()}: Detect results are stored here.
                datum.poseKeypoints {np.ndarray}:
                    shape = [P, N, 3]; type = np.float32.
                    P: Number of detected human bodies. 
                    N: N body joints defined by the detector model, such as COCO and MPI.
                    3: [pixel_column, pixel_row, confidence]. 
                        The not-detected joint has a [0., 0., 0.] value.
                datum.poseKeypoints {np.ndarray}:
                    When there are no humans detected, 
                    this variable becomes a number of the type np.ndarray???
                datum.handKeypoints {list}:
                    List length is 1 or 2, which are the left hand and the right hand.
                    Each element is a np.ndarray of shape = [P, M, 3]
                    P and 3 have the same meaning as `poseKeypoints`.
                datum.handKeypoints {list}:
                    When there are no humans detected, 
                    this variable becomes a list with two numbers???
            body_joints {np.ndarray}:
                 [P, N, 3]. P is humans number; N is joints number.
            body_joints {list or np.ndarray}:
                If params["hand"] == False, return [None]*number_of_people
                Else if no human, return [];
                Else, return np.ndarray of dimension [P, 2, N, 3]

        For model definition, please see:
            https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/output.md
        '''
        datum = op.Datum()
        datum.cvInputData = color_image
        self._opWrapper.emplaceAndPop([datum])
        self._datum = datum
        # print("Body keypoints: \n" + str(datum.poseKeypoints))
        # print("Hand keypoints: \n" + str(datum.handKeypoints))
        if is_return_joints:
            body_joints = datum.poseKeypoints
            if len(body_joints.shape) == 0:  # Detect no people.
                body_joints = []
            number_of_people = len(body_joints)
            body_joints = np.array(body_joints)
            if number_of_people == 0:
                hand_joints = []
            else:
                if self._params["hand"] == False:
                    # The joints of each person is None.
                    hand_joints = [None] * number_of_people
                else:
                    hand_joints = np.array(datum.handKeypoints)
                    hand_joints = hand_joints.swapaxes(
                        0, 1)  # reshape to [P, 2, M, 3]
            return body_joints, hand_joints
        else:
            return datum

    def save_joints_positions(self, datum, pose_filename, hand_filename=None):
        ''' Save body and hand joints to two binary files. 
        For the data type of `datum`, see `def detect`.
        '''

        def save_binary(filename, data):
            if data is not None:
                np.save(filename, np.array(data))

        def save_txt(filename, data):
            if data is not None:
                np.savetxt(filename, np.array(data), delimiter=",")

        save_binary(pose_filename, datum.poseKeypoints)
        save_binary(hand_filename, datum.handKeypoints)
        print(datum.handKeypoints)

        # To load the data, use:
        # bodies_joints = np.load(pose_filename)
        # hands_joints = np.load(hand_filename)

    def set_default_params(self):
        # Refer to $OPENPOSE_HOME/include/openpose/flags.hpp for more parameters.
        params = dict()
        params["model_folder"] = MODEL_PATH
        params["face"] = False  # I haven't done this.
        params["hand"] = True
        params["net_resolution"] = "320x240"  # e.g.: "240x160"
        # params["net_resolution"] = "640x480"  # e.g.: "240x160"
        params["model_pose"] = "COCO"  # Please use "COCO".
        return params

    def set_params(self,
                   command_line_args=[],
                   keys_params={}):
        ''' Set openpose params in place.
        Arguments:
            command_line_args {list}:
                e.g.: ["--hand", False, "--model_pose", "COCO"]
            keys_params {dict}:
                e.g.: {"hand": False, "model_pose": "COCO"}
        '''
        if not isinstance(command_line_args, list):
            raise RuntimeError("`command_line_args` should be list!")
        if not isinstance(keys_params, dict):
            raise RuntimeError("`keys_params` should be dict!")

        params = self._params

        if command_line_args:
            args = command_line_args
            for i in range(0, len(args[1])):
                curr_item = args[1][i]
                if i != len(args[1])-1:
                    next_item = args[1][i+1]
                else:
                    next_item = "1"
                if "--" in curr_item and "--" in next_item:
                    key = curr_item.replace('-', '')
                    if key not in params:
                        params[key] = "1"
                elif "--" in curr_item and "--" not in next_item:
                    key = curr_item.replace('-', '')
                    if key not in params:
                        params[key] = next_item

        if keys_params:
            for key, param in keys_params.iteritems():
                params[key] = param

        return params
    
    def get_img_viz(self):
        return self._datum.cvOutputData

def makedir(folder):
    folder = os.path.dirname(folder)
    if not os.path.exists(folder):
        os.makedirs(folder)


''' -------------------------------------- Unit Test -------------------------------------- '''


def test_openpose_on_images():
    # -- Settings.
    DST_FOLDER = ROOT + "output/"

    # -- Command line arguments.
    def parse_args():
        parser = argparse.ArgumentParser()
        parser.add_argument("--image_dir",
                            default=ROOT+"data/image_i1/color/",
                            help="Process a directory of images. "
                            "Read all standard formats (jpg, png, bmp, etc.).")
        args = parser.parse_known_args()
        return args

    args = parse_args()

    # -- Setup detector.
    detector = OpenposeDetector(
        # keys_params={"hand": False}
    )

    # -- Output folder.
    makedir(DST_FOLDER)

    # -- Read images and detect.
    imagePaths = op.get_images_on_directory(args[0].image_dir)
    for i, imagePath in enumerate(imagePaths):
        t0 = time.time()

        # Read image.
        color_image = cv2.imread(imagePath)

        # Detect human skeletons.
        datum = detector.detect(color_image)
        print("Image {}/{}. Total time = {} seconds.".format(
            i+1, len(imagePaths), time.time()-t0))

        # Save.
        s = DST_FOLDER + "{:05d}".format(i) + "_"
        filename_body = s + "body_joints.npy"
        filename_hand = s + "hand_joints.npy"
        filename_image = s + "result.jpg"

        detector.save_joints_positions(
            datum, filename_body, filename_hand)

        image_with_skeletons_on_it = datum.cvOutputData
        cv2.imwrite(filename_image, image_with_skeletons_on_it)
        print("  Write results to: " + s + "***")

        # Show.
        cv2.imshow("Detection Result", image_with_skeletons_on_it)
        key = cv2.waitKey(15)
        if key == 27:
            break


if __name__ == '__main__':
    test_openpose_on_images()