def __analize_frames_differences__(self, video_path, max_frames=50):
frames, imgs_paths = self.__load_video_frames__(video_path, as_grey=False)
#Original Dynamic Image
if len(frames) > max_frames:
frames2di = frames[0:max_frames]
else:
frames2di = frames
imgPIL, img = getDynamicImage(frames2di)
imgPIL = imgPIL.convert("RGB")
img = np.array(imgPIL)
cv2.imshow("Raw Dynamic Image", img)
key = cv2.waitKey(0)
candidate_frames, frames_indexes = self.__extract_candidate_frames_fromFramesList__(frames)
#Dynamic Image from keyframes
imgPIL, img = getDynamicImage(candidate_frames)
imgPIL = imgPIL.convert("RGB")
img = np.array(imgPIL)
cv2.imshow("Dynamic Image from Keyframes", img)
key = cv2.waitKey(0)
# for i, cframe in enumerate(candidate_frames):
# text = imgs_paths[frames_indexes[i]]
# # show the image
# cv2.putText(cframe, "{}: {:.2f}".format(text, i), (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 0, 255), 3)
# cv2.imshow("Image", cframe)
# key = cv2.waitKey(0)
return frames, candidate_frames, frames_indexes
def __getitem_blur__(self, vid_name):
dynamicImages = []
nelem = 1
sequence = os.listdir(vid_name)
sequence.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
frames, paths = self.loadFramesSeq(vid_name, sequence)
_, indices_segments = self.getVideoSegments(vid_name, idx)
# print(len(frames),len(indices_segments))
# ca2@i@VnTCC5TDK
segment_idxs = []
for i, idxs in enumerate(indices_segments):
frames_in_segment = list(
itemgetter(*idxs)(frames)) #np.asarray(frames)[idxs].tolist()
blurrings = self.extractor.__compute_frames_blurring_fromList__(
frames_in_segment, plot=False)
indexes_candidates = self.extractor.__candidate_frames_blur_based__(
frames_in_segment, blurrings, self.useKeyframes, nelem=nelem)
if nelem > 1:
indexes_candidates = list(
itemgetter(*indexes_candidates)(idxs))
segment_idxs += indexes_candidates
else:
indexes_candidates = idxs[indexes_candidates[0]]
segment_idxs.append(indexes_candidates)
segment_idxs.sort()
candidate_frames = list(itemgetter(*segment_idxs)(frames))
imgPIL, img = dynamicImage.getDynamicImage(candidate_frames)
imgPIL = self.spatial_transform(imgPIL.convert("RGB"))
dynamicImages.append(imgPIL)
return dynamicImages
def __getitem_diff__(self, vid_name):
dynamicImages = []
sequence = os.listdir(vid_name)
sequence.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
video_segment, paths = self.loadFramesSeq(vid_name, sequence)
candidate_frames, frames_indexes = self.extractor.__extract_candidate_frames_fromFramesList__(
video_segment)
# print('Frames indexes-{}/{}='.format(len(sequence),len(frames_indexes)))
imgPIL, img = dynamicImage.getDynamicImage(candidate_frames)
imgPIL = self.spatial_transform(imgPIL.convert("RGB"))
dynamicImages.append(imgPIL)
return dynamicImages
def __avg_dynamic_image__(self, frames, tempWindowLen):
dimages = []
indices = [x for x in range(0, len(frames), 1)]
indices_segments = [indices[x:x + tempWindowLen] for x in range(0, len(indices), tempWindowLen)]
for i, indices_segment in enumerate(indices_segments): #Generate segments using indices
segment = np.asarray(frames)[indices_segment].tolist()
imgPIL, img = getDynamicImage(segment)
dimg = self.__format_dynamic_image__(imgPIL)
dimg = cv2.cvtColor(dimg, cv2.COLOR_RGB2BGR)
dimages.append(dimg)
# avg = np.sum(np.array(dimages), axis=0)
avg=0
return dimages, avg
def __getitem__(self, idx):
vid_name = self.videos[idx]
# print(vid_name, self.numFrames[idx])
label = self.labels[idx]
dynamicImages = []
ipts = []
# preprocessing_time = 0.0
video_segments, _ = self.getVideoSegments(
vid_name, idx
) # bbox_segments: (1, 16, 6)= (no segments,no frames segment,info
paths = []
rgb_central_frames = []
for i, sequence in enumerate(video_segments):
video_segments[i], pths = self.loadFramesSeq(vid_name, sequence)
paths.append(pths)
for sequence in video_segments:
# start_time = time.time()
imgPIL, img = dynamicImage.getDynamicImage(sequence)
rgb = sequence[int(len(sequence) / 2)]
rgb_pil = Image.fromarray(np.uint8(rgb)).convert('RGB')
rgb_central_frames.append(self.rgb_transform(rgb_pil))
# end_time = time.time()
# preprocessing_time += (end_time - start_time)
dynamicImages.append(np.array(imgPIL))
ipts.append(self.spatial_transform(imgPIL.convert("RGB")))
ipts = torch.stack(ipts, dim=0) #torch.Size([bs, ndi, ch, h, w])
rgb_central_frames = torch.stack(rgb_central_frames, dim=0)
# m_bboxes = bbox_from_di(dynamicImages, num_imgs=5, plot=False)
m_bboxes = [0]
gt_bboxes, one_box = None, None
if self.dataset == 'ucfcrime2local':
gt_bboxes, one_box = load_bbox_gt(vid_name, label, paths[0])
else:
one_box = m_bboxes
# print('m_bboxes: ', len(one_box))
# gt_bboxes, one_box = [-1, -1, -1, -1], [0, 0, 224, 224]
one_box = torch.from_numpy(np.array(one_box)).float()
return (ipts, idx, dynamicImages, one_box, rgb_central_frames), label
def main():
extractor = FrameExtractor(len_window=5)
# datasetAll = [os.path.join(constants.PATH_HOCKEY_FRAMES_VIOLENCE, '24'),
# os.path.join(constants.PATH_VIF_FRAMES, '1', 'Violence', 'crowd_violence__Man_Utd_vs_Roma_Crowd_Trouble__uncychris__ZGI5vlDMpJA'),
# os.path.join(constants.PATH_RWF_2000_FRAMES, 'train', 'Fight', '_2RYnSFPD_U_0')]
# datasetAll = [os.path.join(constants.PATH_RWF_2000_FRAMES, 'train', 'Fight', '_2RYnSFPD_U_0'),
# os.path.join(constants.PATH_RWF_2000_FRAMES, 'train', 'Fight', '_q5Nwh4Z6ao_1'),
# os.path.join(constants.PATH_RWF_2000_FRAMES, 'train', 'NonFight','i2sLegg2JPA_1')]
# images_folder = os.path.join(constants.PATH_RWF_2000_FRAMES, 'train', 'Fight','GafFu4IZtIA_0')
# datasetAll, labelsAll, numFramesAll = hockeyLoadData(shuffle=True)
# datasetAll, labelsAll, numFramesAll, splitsLen = vifLoadData(constants.PATH_VIF_FRAMES)
train_names, train_labels, train_num_frames, test_names, test_labels, test_num_frames = rwf_load_data()
datasetAll = train_names + test_names
bb = []
nelem = 1
seqLen = 15
dynamicImgLen = [5, 10, 15, 20]
for k, video_path in enumerate(datasetAll):
print()
_, video_name = os.path.split(video_path)
frames, imgs_paths = extractor.__load_video_frames__(video_path) # Load all frames
print(len(imgs_paths), video_path)
indices_segments = extractor.__getVideoSegments__(video_path, seqLen=seqLen) # split all video in shots of seqLen frames
# print(len(video_segments), video_segments)
print(indices_segments)
segment_idxs = []
for i,idxs in enumerate(indices_segments):
frames_in_segment = list(itemgetter(*idxs)(frames)) #np.asarray(frames)[idxs].tolist()
blurrings = extractor.__compute_frames_blurring_fromList__(frames_in_segment, plot=False)
# print('blurrings=',blurrings)
indexes_candidates = extractor.__candidate_frames_blur_based__(frames_in_segment, blurrings, 'blur-min', nelem=nelem)
# print('indexes_candidates1=',indexes_candidates, type(indexes_candidates))
if nelem > 1:
indexes_candidates = list(itemgetter(*indexes_candidates)(idxs))
segment_idxs += indexes_candidates
print('Segment i{}-(len={}/candidates={})'.format(i+1,len(frames_in_segment), len(indexes_candidates)))
else:
indexes_candidates = idxs[indexes_candidates[0]]
segment_idxs.append(indexes_candidates)
print('Segment i{}-(len={}/candidates={})'.format(i+1,len(frames_in_segment), 1))
print('indexes_candidates=',indexes_candidates)
segment_idxs.sort()
candidate_frames = list(itemgetter(*segment_idxs)(frames))
print('CAndidate frames len({}), indexes={}'.format(len(candidate_frames),segment_idxs))
dyn_image, _ = getDynamicImage(frames[0:30]) #all frames
dyn_image = extractor.__format_dynamic_image__(dyn_image)
cv2.imshow("dyn_image", dyn_image)
key = cv2.waitKey(0)
dyn_image_cand, _ = getDynamicImage(candidate_frames) #candidate frames
dyn_image_cand = extractor.__format_dynamic_image__(dyn_image_cand)
cv2.imshow("dyn_image_cand", dyn_image_cand)
key = cv2.waitKey(0)
def main():
segment_len = 5
splits = [0, 5, 10, 15, 20, 30]
video_segments = []
# segment, central_frame = load_video_segment(os.path.join(constants.PATH_VIF_FRAMES, '1/Violence','fans_violence__Slovak_chauvinism_nationalism_extremism_intolerance_and_all_form'), init_frame=0, num_frames=segment_len)
segment, central_frame = load_video_segment(os.path.join(
constants.PATH_VIF_FRAMES, '1/NonViolence',
'peaceful_football_crowds__MaximsNewsNetwork_SUDAN_ELECTIONS_YASIR_ARMAN_OMAR_AL_BASHIR_UNM'
),
init_frame=0,
num_frames=segment_len)
# segment, central_frame = load_video_segment(os.path.join(constants.PATH_HOCKEY_FRAMES_VIOLENCE, '127'), init_frame=0, num_frames=segment_len)
# segment, central_frame = load_video_segment(os.path.join(constants.PATH_UCFCRIME2LOCAL_FRAMES_VIOLENCE, 'Arrest002-VSplit1'), init_frame=10, num_frames=segment_len)
print('Segment Len=', len(segment))
imgPIL, img = getDynamicImage(segment)
imcv = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
gray = cv2.cvtColor(imcv, cv2.COLOR_BGR2GRAY)
print('Image shape=', gray.shape) #(288, 360)
# Calculate gradient RGB
# sobelx = cv2.Sobel(imcv, cv2.CV_64F, 1, 0, ksize=5)
# sobely = cv2.Sobel(imcv, cv2.CV_64F, 0, 1, ksize=5)
# mag, angle = cv2.cartToPolar(sobelx, sobelx, angleInDegrees=True)
# print('Magnitude: ',mag.shape, '-Angles: ', angle.shape)
pixels_per_cell = (8, 8)
n_orientations = 128
cells_per_block = (1, 1)
fd_rgb, hog_image_rgb = hog(imcv,
orientations=n_orientations,
pixels_per_cell=pixels_per_cell,
cells_per_block=cells_per_block,
visualize=True,
multichannel=True,
feature_vector=False)
fd_gray, hog_image_gray = hog(gray,
orientations=n_orientations,
pixels_per_cell=pixels_per_cell,
cells_per_block=cells_per_block,
visualize=True,
multichannel=False,
feature_vector=False)
print('**hog_image_rgb: ', hog_image_rgb.shape, '-hog_image_gray: ',
hog_image_gray.shape)
print('****descriptor_rgb: ', fd_rgb.shape, '-descriptor_gray: ',
fd_gray.shape)
rows = 1
cols = 3
# plt.subplot(rows,cols,1),plt.imshow(imcv,cmap = 'gray')
# plt.title('Original'), plt.xticks([]), plt.yticks([])
# plt.subplot(rows,cols,2),plt.imshow(sobelx,cmap = 'gray')
# plt.title('Sobel X'), plt.xticks([]), plt.yticks([])
# plt.subplot(rows,cols,3),plt.imshow(sobely,cmap = 'gray')
# plt.title('Sobel Y'), plt.xticks([]), plt.yticks([])
# plt.subplot(rows,cols,4),plt.imshow(hog_image_rgb,cmap = 'gray')
# plt.title('SKImg HOG'), plt.xticks([]), plt.yticks([])
# gray_sobelx8u = cv2.Sobel(gray, cv2.CV_8U, 1, 0, ksize=5)
# gray_sobely8u = cv2.Sobel(gray, cv2.CV_8U, 0, 1, ksize=5)
# plt.subplot(rows,cols,5),plt.imshow(gray,cmap = 'gray')
# plt.title('Original_Gray'), plt.xticks([]), plt.yticks([])
# plt.subplot(rows,cols,6),plt.imshow(gray_sobelx8u,cmap = 'gray')
# plt.title('Sobel CV_8U x'), plt.xticks([]), plt.yticks([])
# plt.subplot(rows,cols,7),plt.imshow(gray_sobely8u,cmap = 'gray')
# plt.title('Sobel CV_8U y'), plt.xticks([]), plt.yticks([])
# Calculate gradient GRAY
gray_sobelx64 = cv2.Sobel(gray, cv2.CV_64F, 1, 0, ksize=5)
gray_sobely64 = cv2.Sobel(gray, cv2.CV_64F, 0, 1, ksize=5)
magnitudes_gray, orientations_gray = cv2.cartToPolar(gray_sobelx64,
gray_sobely64,
angleInDegrees=True)
print('Magnitude_gray: ', magnitudes_gray.shape, '-Angles_gray: ',
orientations_gray.shape)
print('Angles. Max={}, Min={}'.format(np.amax(orientations_gray),
np.amin(orientations_gray)))
plt.subplot(rows, cols, 1), plt.imshow(gray, cmap='gray')
plt.title('Original_Gray'), plt.xticks([]), plt.yticks([])
plt.subplot(rows, cols, 2), plt.imshow(central_frame)
plt.title('Frame'), plt.xticks([]), plt.yticks([])
# plt.subplot(rows,cols,6),plt.imshow(gray_sobelx64,cmap = 'gray')
# plt.title('Sobel CV_64F x'), plt.xticks([]), plt.yticks([])
# plt.subplot(rows,cols,7),plt.imshow(gray_sobely64,cmap = 'gray')
# plt.title('Sobel CV_64F y'), plt.xticks([]), plt.yticks([])
plt.subplot(rows, cols, 3), plt.imshow(hog_image_gray, cmap='gray')
plt.title('SKImg HOG'), plt.xticks([]), plt.yticks([])
plt.show()
entropies = compute_entropy(fd_gray, entropy_type='own', threshold=0)
heat_map = sb.heatmap(entropies, annot=False)
plt.title('Entropy')
plt.show()