def play(self, is_train, video_index):
self.scd = PgdSkeleton(Path.home() / 'intentlong', is_train,
self.img_size)
res = self.scd[video_index]
coord_norm = res[PG.COORD_NORM]
coord_norm = np.transpose(coord_norm, (0, 2, 1))
coord = coord_norm * np.array(self.img_size)
img_shape = self.img_size[::-1] + (3, )
kps = [
KeypointsOnImage.from_xy_array(coord_JX, shape=img_shape)
for coord_JX in coord
]
cap = cv2.VideoCapture(str(res[PG.VIDEO_PATH]))
v_size = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
v_fps = int(cap.get(cv2.CAP_PROP_FPS))
duration = int(1000 / v_fps)
for n in range(v_size):
ret, img = cap.read()
re_img = cv2.resize(img, self.img_size)
pOnImg = kps[n]
img_kps = pOnImg.draw_on_image(re_img)
if self.is_unittest:
break
cv2.imshow("Play saved keypoint results", img_kps)
cv2.waitKey(duration)
cap.release()
def play_dataset_video(self, is_train, video_index, show=True):
self.scd = PgdSkeleton(Path.home() / 'PoliceGestureLong', is_train, self.img_size)
res = self.scd[video_index]
print('Playing %s' % res[PG.VIDEO_NAME])
coord_norm_FXJ = res[PG.COORD_NORM] # Shape: F,X,J
coord_norm_FJX = np.transpose(coord_norm_FXJ, (0, 2, 1)) # FJX
coord = coord_norm_FJX * np.array(self.img_size)
img_shape = self.img_size[::-1] + (3,)
kps = [KeypointsOnImage.from_xy_array(coord_JX, shape=img_shape) for coord_JX in coord] # (frames, KeyOnImage)
cap = cv2.VideoCapture(str(res[PG.VIDEO_PATH]))
v_size = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
v_fps = int(cap.get(cv2.CAP_PROP_FPS))
duration = int(1000/(v_fps*4))
gestures = [] # Full video gesture recognition results
for n in range(v_size):
gdict = self.gpred.from_skeleton(coord_norm_FXJ[n][np.newaxis])
gesture = gdict[PG.OUT_ARGMAX]
gestures.append(gesture)
if not show:
continue
ret, img = cap.read()
re_img = cv2.resize(img, self.img_size)
ges_name = self.gesture_dict[gesture]
re_img = draw_text(re_img, 50, 100, ges_name, (255, 50, 50), size=40)
pOnImg = kps[n]
img_kps = pOnImg.draw_on_image(re_img)
if self.is_unittest:
break
cv2.imshow("Play saved keypoint results", img_kps)
cv2.waitKey(duration)
cap.release()
gestures = np.array(gestures, np.int)
res[PG.PRED_GESTURES] = gestures
print('The prediction of video ', res[PG.VIDEO_NAME], ' is completed')
return res
def save():
ds = PgdSkeleton(Path.home() / 'PoliceGestureLong',
is_train=True,
resize_img_size=(512, 512))
loader = DataLoader(ds,
batch_size=1,
shuffle=False,
num_workers=0,
collate_fn=lambda x: x)
for d in loader:
pass