def main():
args = parse_args()
reset_config(config, args)
image_path = '/home/xyliu/Pictures/pose/soccer.png'
########## 加载human detecotor model
# from lib.detector.yolo.human_detector2 import load_model as yolo_model
# human_model = yolo_model()
from lib.detector.mmdetection.high_api import load_model as mm_model
human_model = mm_model()
# from lib.detector.yolo.human_detector2 import human_bbox_get as yolo_det
from lib.detector.mmdetection.high_api import human_boxes_get as mm_det
# load MODEL
pose_model = model_load(config)
pose_model.eval()
pose_model.cuda()
from pycocotools.coco import COCO
annFile = '/ssd/xyliu/data/coco/annotations/instances_val2017.json'
im_root = '/ssd/xyliu/data/coco/images/val2017/'
coco = COCO(annFile)
catIds = coco.getCatIds(catNms=['person'])
# 所有人体图片的id
imgIds = coco.getImgIds(catIds=catIds)
kpts_result = []
detected_image_num = 0
box_num = 0
for imgId in tqdm(imgIds):
img = coco.loadImgs(imgId)[0]
im_name = img['file_name']
img = im_root + im_name
img_input = cv2.imread(
img, cv2.IMREAD_COLOR | cv2.IMREAD_IGNORE_ORIENTATION)
try:
# bboxs, scores = yolo_det(img_input, human_model)
bboxs, scores = mm_det(human_model, img_input, 0.1)
inputs, origin_img, center, scale = PreProcess(
img_input, bboxs, scores, config)
except Exception as e:
print(e)
continue
detected_image_num += 1
with torch.no_grad():
output = pose_model(inputs.cuda())
preds, maxvals = get_final_preds(config,
output.clone().cpu().numpy(),
np.asarray(center),
np.asarray(scale))
# vis = np.ones(shape=maxvals.shape,)
vis = maxvals
preds = preds.astype(np.float16)
keypoints = np.concatenate((preds, vis), -1)
for k, s in zip(keypoints, scores.tolist()):
box_num += 1
k = k.flatten().tolist()
item = {
"image_id": imgId,
"category_id": 1,
"keypoints": k,
"score": s
}
kpts_result.append(item)
num_joints = 17
in_vis_thre = 0.2
oks_thre = 0.5
oks_nmsed_kpts = []
for i in range(len(kpts_result)):
img_kpts = kpts_result[i]['keypoints']
kpt = np.array(img_kpts).reshape(17, 3)
box_score = kpts_result[i]['score']
kpt_score = 0
valid_num = 0
# each joint for bbox
for n_jt in range(0, num_joints):
# score
t_s = kpt[n_jt][2]
if t_s > in_vis_thre:
kpt_score = kpt_score + t_s
valid_num = valid_num + 1
if valid_num != 0:
kpt_score = kpt_score / valid_num
# rescoring 关节点的置信度 与 box的置信度的乘积
kpts_result[i]['score'] = kpt_score * box_score
import json
data = json.dumps(kpts_result)
print(
'image num is {} \tdetected_image num is {}\t person num is {}'.format(
len(imgIds), detected_image_num, box_num)),
# data = json.dumps(str(kpts_result))
with open('person_keypoints.json', 'wt') as f:
# pass
f.write(data)
def ckpt_time(t0=None, display=None):
if not t0:
return time.time()
else:
t1 = time.time()
if display:
print('consume {:2f} second'.format(t1 - t0))
return t1 - t0, t1
###### 加载human detecotor model
from lib.detector.yolo.human_detector import load_model as yolo_model
# human_model = yolo_model()
from lib.detector.mmdetection.high_api import load_model as mm_model
human_model = mm_model()
def main():
args = parse_args()
update_config(cfg, args)
if not args.camera:
# handle video
cam = cv2.VideoCapture(args.video_input)
video_length = int(cam.get(cv2.CAP_PROP_FRAME_COUNT))
else:
cam = cv2.VideoCapture(0)
video_length = 30000
ret_val, input_image = cam.read()