def face_detection(self, img, scale_step=20):
# this training accuracy should be the same as your training process,
##################################################################################
train_predicts = []
for idx in range(self.data.shape[0]):
train_predicts.append(self.sc_function(self.data[idx, ...]))
print('Check training accuracy is: ',
np.mean(np.sign(train_predicts) == self.labels))
##################################################################################
scales = 1 / np.linspace(1, 8, scale_step)
patches, patch_xyxy = image2patches(scales, img)
print('Face Detection in Progress ..., total %d patches' %
patches.shape[0])
predicts = [self.sc_function(patch) for patch in tqdm(patches)]
print(np.mean(np.array(predicts) > 0), np.sum(np.array(predicts) > 0))
pos_predicts_xyxy = np.array([
patch_xyxy[idx] + [score] for idx, score in enumerate(predicts)
if score > 0
])
if pos_predicts_xyxy.shape[0] == 0:
return
xyxy_after_nms = nms(pos_predicts_xyxy, 0.01)
print('after nms:', xyxy_after_nms.shape[0])
for idx in range(xyxy_after_nms.shape[0]):
pred = xyxy_after_nms[idx, :]
cv2.rectangle(img, (int(pred[0]), int(pred[1])),
(int(pred[2]), int(pred[3])), (0, 255, 0),
2) #gree rectangular with line width 3
return img
def get_hard_negative_patches(self, img, name, scale_step = 10):
scales = 1 / np.linspace(1, 8, scale_step)
if os.path.exists('patches_%s.npy' %name):
patches = np.load('patches_%s.npy' %name)
patch_xyxy = np.load('patch_position_%s.npy' %name)
print('Patches loaded')
else:
patches, patch_xyxy = image2patches(scales, img)
print('Face Detection in Progress ..., total %d patches' % patches.shape[0])
np.save('patches_%s.npy' %name, patches)
np.save('patch_position_%s.npy' %name, patch_xyxy)
print('Patches saved')
if os.path.exists('patches_score%s.pkl' %name):
print('[Find cached Patches Scores, patches_score%s.pkl loading...]' % name)
with open('patches_score%s.pkl' %name, 'rb') as f:
predicts = pickle.load(f)
print('Patches Scores loaded')
else:
predicts = [self.sc_function(patch) for patch in tqdm(patches)]
pickle.dump(predicts, open('patches_score%s.pkl' %name, 'wb'))
print('Patches Scores saved')
predicts = np.array(predicts)
wrong_patches = patches[np.where(predicts > 0), ...]
return wrong_patches
def get_hard_negative_patches(self, img, scale_step = 10):
scales = 1 / np.linspace(1, 8, scale_step)
patches, patch_xyxy = image2patches(scales, img)
print('Get Hard Negative in Progress ..., total %d patches' % patches.shape[0])
predicts = [self.sc_function(patch) for patch in tqdm(patches)]
wrong_patches = patches[np.where(np.array(predicts) > 0), ...]
print(wrong_patches.shape)
return wrong_patches
def get_hard_negative_patches(self, img, scale_step=10):
scales = 1 / np.linspace(1, 8, scale_step)
patches, patch_xyxy = image2patches(scales, img)
print('Get Hard Negative in Progress ..., total %d patches' %
patches.shape[0])
predicts = [self.sc_function(patch) for patch in tqdm(patches)]
predicts = np.array(predicts)
wrong_patches = patches[np.where(predicts > 0), ...]
pos_predicts_xyxy = np.array([
patch_xyxy[idx] + [score] for idx, score in enumerate(predicts)
if score > 0
])
patches_after_nms, nms_patches_idx = nms(pos_predicts_xyxy, 0.01)
return wrong_patches[0], wrong_patches[0][nms_patches_idx]
def get_hard_negative_patches(self, img, scale_step=10):
scales = 1 / np.linspace(1, 8, scale_step)
patches, patches_xyxy = image2patches(scales, img)
print('Get Hard Negative in Progress..., total %d patches' %
patches.shape[0])
preds = np.array([self.sc_function(patch) for patch in tqdm(patches)])
wrong_patches = patches[np.where(preds > 0), ...]
# apply nms
# hard_negatives_xyxy = np.array([patches_xyxy[i] + [score] for i, score in enumerate(preds) if score > 0])
# print('num hard negative detections before nms: ', hard_negatives_xyxy.shape[0])
# hard_negatives_xyxy = nms(hard_negatives_xyxy, 0.01)
# print('num hard negative detections after nms: ', hard_negatives_xyxy.shape[0])
# wrong_patches = np.delete(hard_negatives_xyxy, 4, 1)
return wrong_patches
def face_detection(self, img, name, scale_step = 20):
# this training accuracy should be the same as your training process,
##################################################################################
#train_predicts = []
#for idx in range(self.data.shape[0]):
# train_predicts.append(self.sc_function(self.data[idx, ...]))
#print('Check training accuracy is: ', np.mean(np.sign(train_predicts) == self.labels))
##################################################################################
print(scale_step)
scales = 1 / np.linspace(1, 8, scale_step)
if os.path.exists('patches_%s.npy' %name):
patches = np.load('patches_%s.npy' %name)
patch_xyxy = np.load('patch_position_%s.npy' %name)
print('Patches loaded')
else:
patches, patch_xyxy = image2patches(scales, img)
print('Face Detection in Progress ..., total %d patches' % patches.shape[0])
np.save('patches_%s.npy' %name, patches)
np.save('patch_position_%s.npy' %name, patch_xyxy)
print('Patches saved')
if os.path.exists('patches_score%s.pkl' %name):
print('[Find cached Patches Scores, patches_score%s.pkl loading...]' % name)
with open('patches_score%s.pkl' %name, 'rb') as f:
predicts = pickle.load(f)
print('Patches Scores loaded')
else:
predicts = [self.sc_function(patch) for patch in tqdm(patches)]
pickle.dump(predicts, open('patches_score%s.pkl' %name, 'wb'))
print('Patches Scores saved')
print(np.mean(np.array(predicts) > 0), np.sum(np.array(predicts) > 0))
pos_predicts_xyxy = np.array([np.hstack((patch_xyxy[idx], np.array(score))) for idx, score in enumerate(predicts) if score > 0])
if pos_predicts_xyxy.shape[0] == 0:
return
xyxy_after_nms = nms(pos_predicts_xyxy, 0.01)
print('after nms:', xyxy_after_nms.shape[0],xyxy_after_nms.shape[1])
for idx in range(xyxy_after_nms.shape[0]):
pred = xyxy_after_nms[idx, :]
cv2.rectangle(img, (int(pred[0]), int(pred[1])), (int(pred[2]), int(pred[3])), (0, 255, 0), 2) #gree rectangular with line width 3
return img
def face_detection(self, img, scale_step=20):
train_preds = []
for i in range(self.data.shape[0]):
train_preds.append(self.sc_function(self.data[i, ...]))
tr_acc = np.mean(np.sign(train_preds) == self.labels)
print('Check on training accuracy is: ', tr_acc)
pp = 'pos_preds_xyxy_face_1_with_negatives.pkl' # CHANGE FILE
if os.path.exists(pp):
pos_preds_xyxy = get_data(pp)
else:
scales = 1 / np.linspace(1, 8, scale_step)
patches, patches_xyxy = image2patches(scales, img)
print('Face Detection in Progress..., total %d patches' %
patches.shape[0])
preds = [self.sc_function(patch) for patch in tqdm(patches)]
pos_preds = np.array(preds) > 0
print('positive detections mean and sum: ', np.mean(pos_preds),
np.sum(pos_preds))
pos_preds_xyxy = np.array([
patches_xyxy[i] + [score] for i, score in enumerate(preds)
if score > 0
])
save_data(
pos_preds_xyxy,
'pos_preds_xyxy_face_1_with_negatives.pkl') # CHANGE FILE
if pos_preds_xyxy.shape[0] == 0:
return
print('num positive detections before nms: ', pos_preds_xyxy.shape[0])
pos_preds_xyxy = nms(pos_preds_xyxy, 0.01)
print('num positive detections after nms: ', pos_preds_xyxy.shape[0])
for i in range(pos_preds_xyxy.shape[0]):
pred = pos_preds_xyxy[i, :]
xy1 = (int(pred[0]), int(pred[1]))
xy2 = (int(pred[2]), int(pred[3]))
cv2.rectangle(img, xy1, xy2, (0, 255, 0), 2)
return img
