def get_pose(image_dir, img_name, pose_dir):
inference = tf_pose.infer(image=os.path.join(image_dir, img_name))[0]
body_parts = inference.body_parts
person_data = []
for i in range(18):
try:
person = body_parts[i]
x = person.x * 192
y = person.y * 256
confidence = person.score
except:
x = 0
y = 0
confidence = 0
data = [x, y, confidence]
person_data += data
output = {}
output['version'] = 1.0
people = dict()
people['face_keypoints'] = []
people['pose_keypoints'] = person_data
people['hand_right_keypoints'] = []
people['hand_left_keypoints'] = []
_people = [people]
output['people'] = _people
with open(
os.path.join(pose_dir, img_name.replace('.jpg',
'_keypoints.json')),
'w') as file:
json.dump(output, file)
def get_pose():
# load video
path_list = ["Sub04_03_30_FS_R11.avi", "Sub05_01_30_FK_R21.avi",
"Sub05_02_30_KS_R11.avi", "Sub08_04_60_KS_R21.avi"]
for i, img_path in enumerate(path_list):
print(i)
print(img_path)
# dummy
pose_all = []
# video
cap = cv2.VideoCapture(img_path)
while(True):
# Capture frame-by-frame
ret, frame = cap.read()
if ret:
# pose detection
frame = frame[0:480, 50:480, :]
coco_style = tf_pose.infer(frame)
# check
if len(coco_style)<1:
temp = np.ones((34,1))*-1.
else:
temp = np.array(coco_style[0][0])
temp = np.expand_dims(np.delete(temp, range(2,51,3)), axis=1)
temp=temp.tolist()
pose_all.append(temp)
else:
break
# print(i)
np.save("./"+str(img_path[0:-4])+".npy", pose_all)
# When everything done, release the capture
cap.release()
def all_image(self, img_dir="/home/externalDisk/RULA_image/images_lift/side/"):
from PIL import Image
import glob
image_list = []
fname_list = []
pose_list = []
for i, file_path in enumerate(glob.glob(img_dir+"*.png")):
print(i)
## append file name
fname=file_path[47:]
fname_list.append(fname)
print(fname)
## append image
im = np.array(Image.open(file_path))
image_list.append(im)
## append predicted pose
pose = tf_pose.infer(im)
try:
pose = np.array(pose[0][0])
except:
pose = np.zeros((51,))
pose_list.append(pose)
all3=[]
all3.append(fname_list)
all3.append(image_list)
all3.append(pose_list)
print(len(all3))
import pickle
with open("img_pose", "wb") as f:
pickle.dump(all3, f)
def continuous_predict(self, verbose=False):
while (self.predicting):
self.file_helper.update_image_library()
image_path, image_timestamp_name = self.file_helper.get_new_image()
if image_path is not None:
coco_style = tf_pose.infer(image_path)
# coco_style = [ 'asdf', 'asdf2']
self.file_helper.save_misc(coco_style, image_timestamp_name)
if verbose:
print(coco_style)
time.sleep(0.001)
def load_save_h36():
# load pose
import sys
sys.path.append("/home/lli40/PyCode/Downloaded/tf-pose-estimation")
import tf_pose
img_path = "/home/lli40/PyCode/MyProject/RULA_2DImage/data/pose100.npy"
img = np.load(img_path, allow_pickle=True, encoding="latin1")
cnt = 0
pose_all = []
# import time
for i in range(len(img)):
# start_time = time.time()
cnt += 1
frame = img[i] * 255.
frame = frame.astype(np.uint8)
coco_style = tf_pose.infer(frame)
# elapsed_time = time.time() - start_time
# print("FPS: " + str(1. / elapsed_time))
if len(coco_style) < 1:
temp = np.ones((34, 1)) * -1.
else:
temp = np.array(coco_style[0][0])
temp = np.expand_dims(np.delete(temp, range(2, 51, 3)), axis=1)
print(temp.shape)
temp = temp.tolist()
pose_all.append(temp)
exit()
pose_all = np.array(pose_all)
print(pose_all.shape)
# load gscore
gscore_name = np.loadtxt(
"/home/lli40/PyCode/MyProject/RULA_2DImage/data/gscore-name-human36.txt",
dtype=str)
gscore_name = np.delete(gscore_name, 0)
# dummy
gscore_all = []
for i in range(len(gscore_name)):
idx = gscore_name[i][0:-3]
temp = gscore_name[i][-1]
gscore_all.append(int(temp))
# to np array
gscore_all = np.array(gscore_all)
# save
np.save("/home/lli40/PyCode/MyProject/RULA_2DImage/data/x_all_h36_raw.npy",
pose_all)
np.save("/home/lli40/PyCode/MyProject/RULA_2DImage/data/y_all_h36_raw.npy",
gscore_all)
def data_upload_lift():
## for lift
lift_img_dir = "/home/externalDisk/RULA_image/images_lift/side/"
from PIL import Image
import glob
image_list = []
fname_list = []
import cv2
import sys
sys.path.append("/home/lli40/PyCode/Downloaded/tf-pose-estimation")
import tf_pose
for i, file_path in enumerate(glob.glob(lift_img_dir + "*.png")):
print(i)
## append file name
fname = file_path[47:]
fname_list.append(fname)
print(fname)
## append image
im = np.array(Image.open(file_path))
image = im[0:480, 50:480, :]
coco_style = tf_pose.infer(image)
if len(coco_style) > 0:
joints = np.array(coco_style[0][0])
joints = np.round(joints).astype(int)
image = image.copy()
# draw rectangle on face
if joints[0] != 0 or joints[0] != -1:
cv2.rectangle(image, (joints[0] - 30, joints[1] - 30),
(joints[0] + 5, joints[1] + 30), (255, 255, 255),
-1)
image_list.append(image)
# cv2.imshow("sample", image)
# cv2.waitKey(0)
print(np.array(image_list).shape)
np.save(
"/home/lli40/PyCode/MyProject/RULA_2DImage/data_upload/lift_image.npy",
image_list)
def img2img(self):
img_path="/home/lli40/PyCode/MyProject/2DPose_detect/Data/Human80K/GTD2P/x_test.npy"
img = np.load(img_path, allow_pickle=True, encoding="latin1")
cnt=0
for i in range(len(img)):
cnt+=1
frame=img[i]*255.
frame=frame.astype(np.uint8)
coco_style = tf_pose.infer(frame)
if len(coco_style)<1:
continue
joints = np.array(coco_style[0][0])
cv2.imwrite("./human80/" + str(cnt) + ".png", frame)
visualizer = vis()
new_frame = visualizer.vis_pose(image=frame, joints=joints, is_display = False)
cv2.imwrite("./human80/" + str(cnt) + "_n.png", new_frame)
print(cnt)
def video2video(self, source_path, target_path):
cap = cv2.VideoCapture(source_path)
cnt= 0
while (cap.isOpened()):
cnt += 1
ret, frame = cap.read()
frame = frame[0:480, 50:480, :]
coco_style = tf_pose.infer(frame)
print(len(coco_style))
joints = np.array(coco_style[0][0])
visualizer = vis()
new_frame = visualizer.vis_pose(image=frame, joints=joints)
# new_frame = frame
new_frame = new_frame[:, 200:-1, :]
frame = frame [:, 200:-1, :]
cv2.imwrite("./lift/" + str(cnt) + ".png", frame)
cv2.imwrite("./lift/"+str(cnt)+"_n.png", new_frame)
print(cnt)
# cv2.imshow('frame', new_frame)
#
# if cv2.waitKey(0) == ord("q"):
# continue
# else:
# exit()
# When everything done, release the capture
cap.release()
cv2.destroyAllWindows()
def get_video():
# load video
path_list = ["Sub04_01_30_FK_R11.avi", "Sub04_03_30_FS_R11.avi", "Sub04_02_30_KS_R21.avi",
"Sub05_01_30_FK_R21.avi", "Sub05_09_30_FS_R11.avi", "Sub05_02_30_KS_R11.avi",
"Sub08_01_00_FK_R11.avi,", "Sub08_11_30_FS_R11.avi", "Sub08_04_60_KS_R21.avi"]
path_list=[ "Sub08_01_00_FS_R11.avi"]
# load gscore
gscore_name = np.loadtxt("/home/lli40/PyCode/MyProject/RULA_2DImage/data/gscore-name-lift.txt", dtype=str)
gscore_name = np.delete(gscore_name, 0)
for i, video_path in enumerate(path_list):
print(video_path)
cap = cv2.VideoCapture(video_path)
fid = 0
fourcc = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter(video_path[0:-4]+"_p.avi", fourcc, 20.0, (1440, 480))
# cv2.namedWindow("output", cv2.WINDOW_NORMAL)
ret = True
while ret:
# Capture frame-by-frame
ret, frame = cap.read()
if ret:
print(str(fid)+" th frame")
fid+=1
# pose detection
# frame = frame[0:480, 50:480, :]
coco_style = tf_pose.infer(frame)
# check
if len(coco_style)<1:
temp = np.ones((34,1))*-1.
else:
temp = np.array(coco_style[0][0])
print(temp.shape)
visualizer = vis()
new_frame = visualizer.vis_pose(image=frame, joints=temp)
print(new_frame.shape)
# cv2.imshow("a",new_frame)
# cv2.waitKey(0)
# exit()
out.write(new_frame)
# cv2.imshow("output", new_frame)
# cv2.waitKey(0)
cap.release()
out.release()
def h36_video():
img = np.load("/home/lli40/PyCode/MyProject/RULA_2DImage/data/pose100.npy", allow_pickle=True, encoding="latin1")
fourcc = cv2.VideoWriter_fourcc(*'XVID')
out = cv2.VideoWriter("H36_p.avi", fourcc, 30.0, (200, 150))
for i in range(len(img)):
frame = img[i] * 255.
frame = frame.astype(np.uint8)
coco_style = tf_pose.infer(frame)
if len(coco_style) < 1:
temp = np.ones((34, 1)) * -1.
else:
temp = np.array(coco_style[0][0])
print(temp.shape)
visualizer = vis()
new_frame = visualizer.vis_pose(image=frame, joints=temp)
print(new_frame.shape)
for j in range(100):
out.write(new_frame)
out.release()
handpose = HandPose(input_tensor)
handpose.load_weights('checkpoints/pretrain')
# TEST_IMAGE_PATH = 'hand1.jpg'
# image_orig = cv2.imread(TEST_IMAGE_PATH)
# handpose.predict(image_orig, 1.5, True, True)
import pdb
import matplotlib.pyplot as plt
import math
TEST_IMAGE_PATH = 'hand2.jpg'
image = cv2.imread(TEST_IMAGE_PATH)
img_w = image.shape[1]
img_h = image.shape[0]
bodys = tf_pose.infer('hand2.jpg')
for body in bodys:
# l_roi_image = handpose.getLeftHandRoiImage(body.body_parts, image)
# r_roi_image = handpose.getRightHandRoiImage(body.body_parts, image)
# handpose.fix_predict(l_roi_image)
lhpts, rhpts = handpose.getHandsPart(image, body)
for partid in lhpts:
part = lhpts[partid]
cv2.circle(image, (int(img_w * part.x), int(img_h * part.y)), 10,
(255, 0, 0), 2)
for partid in rhpts:
part = rhpts[partid]
cv2.circle(image, (int(img_w * part.x), int(img_h * part.y)), 10,
(255, 0, 0), 2)
dnames = [d["photo"].split("/")[-1] for d in data]
print("Start loading photos")
for d in tqdm(data):
load(d["photo"])
print("Start preparing photos:", len(done) - len(photos_name))
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
import tf_pose
for i, d in tqdm(enumerate(data)):
if d['photo'] in photos_name:
continue
try:
coco_style = tf_pose.infer(
f"/home/petr/Documents/Projects/Recognition/data/{d['photo']}")
except Exception as e:
print(e, f"data/{d['photo']}")
continue
photos.append({"name": d['photo'], "points": coco_style, "shape": None})
photos_name.append(d['photo'])
if i % 150 == 0:
with open("photos.pk", "wb") as f:
pickle.dump(photos, f)
with open("photos.pk", "wb") as f:
pickle.dump(photos, f)
print("Complete", len(data), "total:", len(photos))