def inference_video(opt):
#change to cpu/cuda
DET = S3FD(device='cuda')
flist = glob.glob(os.path.join(opt.frames_dir,opt.reference,'*.jpg'))
flist.sort()
dets = []
for fidx, fname in enumerate(flist):
start_time = time.time()
image = cv2.imread(fname)
image_np = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
bboxes = DET.detect_faces(image_np, conf_th=0.9, scales=[opt.facedet_scale])
dets.append([]);
for bbox in bboxes:
dets[-1].append({'frame':fidx, 'bbox':(bbox[:-1]).tolist(), 'conf':bbox[-1]})
elapsed_time = time.time() - start_time
print('%s-%05d; %d dets; %.2f Hz' % (os.path.join(opt.avi_dir,opt.reference,'video.avi'),fidx,len(dets[-1]),(1/elapsed_time)))
savepath = os.path.join(opt.work_dir,opt.reference, 'faces.pckl')
with open(savepath, 'wb') as fil:
pickle.dump(dets, fil)
return dets
def inference_video(opt):
DET = S3FD(device="cuda")
flist = glob.glob(os.path.join(opt.frames_dir, opt.reference, "*.jpg"))
flist.sort()
dets = []
for fidx, fname in enumerate(flist):
start_time = time.time()
image = cv2.imread(fname)
image_np = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
bboxes = DET.detect_faces(image_np,
conf_th=0.1,
scales=[opt.facedet_scale])
dets.append([])
max_conf = 0
temp = {"frame": fidx, "bbox": [400, 400, 650, 650], "conf": .1}
for bbox in bboxes:
# Only get the most confident one
if bbox[-1] >= max_conf:
max_conf = bbox[-1]
temp = {
"frame": fidx,
"bbox": (bbox[:-1]).tolist(),
"conf": bbox[-1]
}
dets[-1].append(temp)
elapsed_time = time.time() - start_time
# print(
# "%s-%05d %d dets %.2f Hz"
# % (
# os.path.join(opt.avi_dir, opt.reference, "video.avi"),
# fidx,
# len(dets[-1]),
# (1 / elapsed_time),
# )
# )
savepath = os.path.join(opt.work_dir, opt.reference, "faces.pckl")
with open(savepath, "wb") as fil:
pickle.dump(dets, fil)
return dets
axeslist.ravel()[ind].imshow(figures[title], cmap=plt.gray())
axeslist.ravel()[ind].set_title(title)
plt.tight_layout()
plt.show()
# load image with cv in RGB.
IMAGE_PATH = 'selfie.jpg'
img = cv2.imread(IMAGE_PATH)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# load detectors.
DET1 = MTCNN(device='cuda')
DET2 = FaceBoxes(device='cuda')
DET3 = TinyFace(device='cuda')
DET4 = PyramidBox(device='cuda')
DET5 = S3FD(device='cuda')
DET6 = DSFD(device='cuda')
# MTCNN returns bboxes and landmarks.
t = time.time()
bboxes, _ = DET1.detect_faces(img, conf_th=0.9, scales=[1])
print('MTCNN : %d faces in %.4f seconds.' % (len(bboxes), time.time() - t))
img1 = draw_bboxes(img, bboxes)
sizes = []
for box in bboxes:
sizes.append((box[2] - box[0]) * (box[3] - box[1]))
print(min(sizes))
print(max(sizes))
# FaceBoxes returns bboxes.
t = time.time()
if os.path.exists(os.path.join(opt.avi_dir, reference)):
rmtree(os.path.join(opt.avi_dir, reference))
if os.path.exists(os.path.join(opt.frames_dir, reference)):
rmtree(os.path.join(opt.frames_dir, reference))
except:
pass
# ========== ========== ========== ==========
# # MAIN
# ========== ========== ========== ==========
# ========== LOADING MANDATORY STUFF ==========
DET = S3FD(device='cuda')
s = SyncNetInstanceAnalyze()
s.loadParameters(opt.syncnet_model)
#print("Model %s loaded."%opt.syncnet_model);
# ========== FOR LOOP OVER ALL EXISTING MP4 FILES IN DATA_PATH ==========
videofiles = glob.glob(opt.data_path + '*/*.mp4')
videofiles = [
v for i, v in enumerate(videofiles) if (i % opt.parts_num) == opt.part
]
#print('Started processing for {} GPUs'.format(opt.nthreads))
jobs = [(vfile, opt, i % opt.nthreads) for i, vfile in enumerate(videofiles)]
if opt.nthreads == 1:
def main():
args = get_args()
method = args.method
device = args.device
conf_threshold = args.conf
iou_threshold = args.iou
WD = WIDERDataset()
if method == 'mtcnn':
det = MTCNN(device=device)
scale_list = [0.0625, 0.125, 0.25, 0.5, 1, 2]
elif method == 'tinyface':
det = TinyFace(device=device)
scale_list = [0.5, 1]
elif method == 's3fd':
det = S3FD(device=device)
scale_list = [0.5, 1]
elif method == 'dsfd':
det = DSFD(device=device)
scale_list = [0.5, 1]
elif method == 'faceboxes':
det = FaceBoxes(device=device)
scale_list = [1, 2, 4]
N = len(WD.data_dict.keys())
total_iou = 0.0
total_recall = 0.0
total_precision = 0.0
total_f1score = 0.0
total_time = 0.0
for image_index, image_name in enumerate(WD.data_dict.keys(), 1):
print('%5d / %5d : %s' % (image_index, N, image_name))
image = cv2.imread(os.path.join(IMAGE_DIR, image_name))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
boxes = WD.data_dict[image_name]
if method == 'mtcnn':
img_time = time.time()
pred_boxes, _ = det.detect_faces(image, conf_threshold, scale_list)
img_time = time.time() - img_time
else:
img_time = time.time()
pred_boxes = det.detect_faces(image, conf_threshold, scale_list)
img_time = time.time() - img_time
true_num = len(boxes)
positive_num = len(pred_boxes)
img_iou = 0.0
img_recall = 0.0
img_precision = 0.0
img_f1score = 0.0
pred_dict = dict()
for box in boxes:
max_iou = 0
for i, pred_box in enumerate(pred_boxes):
if i not in pred_dict.keys():
pred_dict[i] = 0
iou = IoU(box, pred_box)
if iou > max_iou:
max_iou = iou
if iou > pred_dict[i]:
pred_dict[i] = iou
img_iou += max_iou
if true_num * positive_num > 0:
true_positive = 0.0
for i in pred_dict.keys():
if pred_dict[i] > iou_threshold:
true_positive += 1.0
img_recall = true_positive / true_num
img_precision = true_positive / positive_num
if img_recall * img_precision == 0:
img_f1score = 0.0
else:
img_f1score = (2*img_recall*img_precision) / (img_recall+img_precision)
img_iou = img_iou / true_num
print('- | TP = %02d | TN = |' % (true_positive))
print(' | FP = %02d | FN = %02d |' % (positive_num - true_positive, true_num - true_positive))
total_iou += img_iou
total_recall += img_recall
total_precision += img_precision
total_f1score += img_f1score
total_time += img_time
print('- Avg. IoU =', total_iou / image_index)
print('- Avg. Recall =', total_recall / image_index)
print('- Avg. Precision =', total_precision / image_index)
print('- Avg. F1-score =', total_f1score / image_index)
print('- Avg. Inference Time =', total_time / image_index)
print()