def process_df_frames_train_set(item, method='surf', histogram_norm='l1'):
import FeatureExtraction
import random
filename, df_frames = item
if df_frames['frame_index'].iloc[0] % 100 != 24:
return []
boxes_dict = Preprocessing.extract_box_img(df_frames)
train_set = []
kmeans = BOW.load_kmeans(f'kmeans_{method}_128.pkl')
for category, boxes in boxes_dict.items():
for box in boxes:
_, des = FeatureExtraction.feature_extraction(box, method)
if des is not None:
histogram = BOW.compute_histogram(des,
kmeans,
method=method,
norm=histogram_norm)
train_set.append((histogram, category))
return train_set
def clfs_rolling_windows(img,
clf,
stride=8,
padding=0,
ratios=((1, 1), (3, 1), (1, 3)),
scale=1.5,
prob_threshold=0.75,
min_side_length=16,
method='sift',
kmeans_path=None):
if method not in ('sift', 'surf'):
raise ValueError("method param should be sift or surf")
if kmeans_path is None:
kmeans_path = f'kmeans_{method}_128.pkl'
h, w = img.shape[:2]
padded_image = np.zeros((h + padding * 2, w + padding * 2, img.shape[2]),
dtype='uint8')
padded_image[padding:h + padding, padding:w + padding, :] = img
h, w = padded_image.shape[:2]
box_hist_dict = {
'box': [],
'histogram': [],
'high_prob': [],
'high_category': []
}
kmeans = BOW.load_kmeans(kmeans_path)
kp, des = FeatureExtraction.feature_extraction(img, method)
df_kp = pd.DataFrame({
'x': [item.pt[0] for item in kp],
'y': [item.pt[1] for item in kp]
})
for ratio in ratios:
print(f'ratio {ratio}')
ratio_h, ratio_w = ratio
if ratio_h < ratio_w:
# (h, w)
start_window = (min_side_length,
int(min_side_length * ratio_w / ratio_h))
else:
start_window = (int(min_side_length * ratio_h / ratio_w),
min_side_length)
window = start_window
this_stride = stride
tmp = window
count = 0
while tmp[0] <= h and tmp[1] <= w:
count += 1
tmp = (int(tmp[0] * scale), int(tmp[1] * scale))
pbar = tqdm(total=count)
while window[0] <= h and window[1] <= w:
for i in range(0, h - window[0], this_stride):
for j in range(0, w - window[1], this_stride):
#print(i, j)
box = (i, j, i + window[0], j + window[1])
#img_this = padded_image[box[0]:box[2], box[1]:box[3]]
#_, des_this = FeatureExtraction.feature_extraction(img_this, method)
des_this = des[np.logical_and.reduce(
(box[0] < df_kp['y'], df_kp['y'] < box[2],
box[1] < df_kp['x'], df_kp['x'] < box[3]))]
if des_this is None or len(des_this) == 0:
continue
histogram = BOW.compute_histogram(des_this,
kmeans,
method=method,
norm='l0')
if histogram is None:
continue
#print(clf.predict(histogram.reshape(1, -1)))
"""cv2.namedWindow('image', cv2.WINDOW_AUTOSIZE)
cv2.imshow('image', img_this)
cv2.waitKey()
cv2.destroyAllWindows()"""
box_hist_dict['box'].append(box)
box_hist_dict['histogram'].append(histogram)
window = (int(window[0] * scale), int(window[1] * scale))
this_stride = int(this_stride * np.sqrt(scale))
pbar.update(1)
pbar.close()
X = np.array(box_hist_dict['histogram'])
y_prob = clf.predict_proba(X)
high_prob = np.max(y_prob, axis=1).ravel().tolist()
high_category = np.argmax(y_prob, axis=1).ravel().tolist()
box_hist_dict['high_prob'] = high_prob
box_hist_dict['high_category'] = high_category
df_box_hist = pd.DataFrame(box_hist_dict)
df_box_hist = df_box_hist[df_box_hist['high_category'] != 0]
df_box_hist = df_box_hist[df_box_hist['high_prob'] > prob_threshold]
return df_box_hist
