def main():
segment_database = SegmentDatabase()
label = 'vertical_oscillation'
segments = segment_database.get_with_label(label)
print '#segments for label %s: %d' % (label, len(segments))
tp = 0
fp = 0
segment_count = 0
for segment in segments:
try:
segment_count += 1
filename = segment.get('ytid')
start_frame, end_frame = segment.get('s')
# print 'showing %s from %d to %d' % (filename, start_frame, end_frame)
print 'segment length: %d' % (end_frame-start_frame)
result, num_frames = show_video(filename, start_frame, end_frame, label)
if result:
tp += num_frames
else:
fp += num_frames
print '%2.2f%% done' % (100.0 * segment_count / float(len(segments)))
except Exception as e:
print e
print 'Label: %s' % label
print 'True positives: %d' % tp
print 'False positives: %d' % fp
def __init__(self, name, ingredients, datasets=[]): self.name = name self.ingredients = ingredients self.segment_database = SegmentDatabase() self.ignore_list = [] self.frames = [] self.result = dict(segments=[]) self.datasets = datasets if datasets: self.segment_database.trim_to_datasets(datasets) self.filename = filename_generator(path='./data/out/')
class Recipe:
def __init__(self, name, ingredients, datasets=[]):
self.name = name
self.ingredients = ingredients
self.segment_database = SegmentDatabase()
self.ignore_list = []
self.frames = []
self.result = dict(segments=[])
self.datasets = datasets
if datasets:
self.segment_database.trim_to_datasets(datasets)
self.filename = filename_generator(path='./data/out/')
def dump_to_json(self):
filename = '%s.json' % self.filename
dump = dict(
ingredients=[ingredient.__to_dict__() for ingredient in self.ingredients],
result=self.result,
filename='%s.avi' % self.filename,
datasets=self.datasets,
)
f = open(filename, 'w')
f.write(json.dumps(dump))
f.close()
def get_frames(self, segment, captions=''):
# print 'segment: ', segment
start, end = segment.get('start'), segment.get('end')
ytid = segment.get('ytid')
cap = get_capture(ytid)
fps = get_fps(cap)
assert(fps == 24)
captions += ' :: %s' % (self.name)
frames = []
if cap.set(cv.CV_CAP_PROP_POS_FRAMES, start):
length = end - start
print 'grabbing %d frames' % length
for i in range(length):
ret, frame = cap.read()
if ret:
# draw_str(frame, (20, 20), '%s: %s' % (ytid, captions))
frames.append(frame)
else:
# problem is that setting a given frame in the video capture may actually jump forward to the closest keyframe
# because of limitations with the video compression algorithm. for short clips this is likely to occur
print 'ERROR: segment read cut short at index %d for segment %s' % (i, json.dumps(segment))
break
# raise Exception('no more frames, index = %d, segment: %s' % (i, json.dumps(segment)))
else:
raise Exception('unable to set seek position in video capture')
print 'returning %d frames' % len(frames)
return frames
def bake(self):
segment_database = self.segment_database
frames = []
for ingredient in self.ingredients:
labels, excl_labels, min_span, max_span = ingredient.labels, [], ingredient.min_span, ingredient.max_span
print 'labels=%s, excl_labels=%s, min_span=%d, max_span=%d' % (labels, excl_labels, min_span, max_span)
candidate_factory = CandidateFactory(segment_database)
candidates = candidate_factory.get_candidates(labels)
# print '#immediate candidates: %d' % len(candidates)
ytids = candidate_factory.get_ytids(20)
query = dict(
labels=labels,
required=ingredient.required_labels,
forbidden=ingredient.forbidden_labels,
)
interval = ingredient.interval
spanAlpha = ingredient.span_alpha
minSpan = ingredient.min_span
maxSpan = ingredient.max_span
ignore_list = self.ignore_list
candidates = sortCandidates(ytids=ytids, query=query, minSpan=minSpan, maxSpan=maxSpan, interval=interval, spanAlpha=spanAlpha, segment_database=segment_database, ignoreList=ignore_list)
# print '#candidates: %d' % len(candidates)
if candidates:
# filter out invalid candidates
candidates = filter(lambda x: x.get('score') >= 0, candidates)
# compute the sum-square of scores
score_sum = sum([candidate.get('score')**2 for candidate in candidates])
if candidates: # if there are any valid candidates left
new_candidates = []
# we want the multiplier to be atleast 1e5
min_score = candidates[-1].get('score')
if min_score:
multiplier = min(1e5, max(1e5, 1 / (min_score / score_sum)))
else:
multiplier = 1e5
# print 'multiplier: ' , multiplier
for candidate in candidates:
# we want atleast 1 occurence
probability_score = multiplier * candidate.get('score')**2 / score_sum
# print 'probability_score: ', probability_score
# print 'rel. rounding error: %2.2f%%' % (100 * abs(probability_score - max(1, int(round(probability_score)))) / probability_score)
new_candidates += [candidate] * max(1, int(round(probability_score)))
# print len(new_candidates)
candidate = random.choice(new_candidates)
print candidate
start = candidate.get('start')
end = candidate.get('end')
ytid = candidate.get('ytid')
score = candidate.get('score')
self.result.get('segments').append(dict(ytid=ytid, start=start, end=end, score=score))
frames += self.get_frames(candidate, captions='score: %2.2f' % (score))
candidate['start'] = max(0, start-10*24)
# no consequence if this actually extends the length of the video
candidate['end'] = end+10*24
self.ignore_list.append(candidate)
else:
print '\n******************\nno candidate matching: %s\n******************\n' % ingredient
else:
print '\n******************\nno candidate matching: %s\n******************\n' % ingredient
self.frames = frames
self.dump_to_json()
return self
def write_video(self, height=360, width=640, fps=24, fourcc=cv.CV_FOURCC('D','I','V','3')):
filename = '%s.avi' % self.filename
frames = self.frames
writer = cv.CreateVideoWriter(filename, int(fourcc),fps,(int(width),int(height)),1)
for frame in frames:
# convert frame to iPlImage
frame = cv.fromarray(frame)
cv_img = cv.CreateImageHeader((width, height), cv.IPL_DEPTH_8U, 3)
cv.SetData(cv_img, frame.tostring(), width * 3)
# write image
cv.WriteFrame(writer, cv_img)
def show_video(self):
fps = 24
for frame in self.frames:
cv2.imshow('', frame)
cv2.waitKey(int(1000/fps))
