def map(self, image_id, image_binary):
try:
boxes = self.image_box_fns[image_id]
except KeyError:
pass
else:
image = imfeat.image_fromstring(image_binary)
size_array = np.array([
image.shape[0], image.shape[1], image.shape[0], image.shape[1]
])
scale_boxes = {}
for box, fn in boxes:
fbox = size_array * box
scale = int(
np.round(np.log2(
(fbox[2] - fbox[0]) / feature.PATCH_SIZE)))
scale_check = int(
np.round(np.log2(
(fbox[3] - fbox[1]) / feature.PATCH_SIZE)))
if scale != scale_check:
raise ValueError('Box is not square.')
scale_boxes.setdefault(scale, []).append((box, fn))
# Order boxes and fn's by scale
for scale in range(max(scale_boxes.keys()) + 1):
if scale > 0:
height, width = np.array(image.shape[:2]) / 2
image = image[:height * 2, :width * 2, :]
if min(width, height) < 1:
raise ValueError('Image is too small')
image = cv2.resize(image, (width, height))
if image is None: # NOTE(brandyn): It is too small
raise ValueError('Image is too small')
try:
boxes = scale_boxes[scale]
except KeyError:
continue
size_array = np.array([
image.shape[0], image.shape[1], image.shape[0],
image.shape[1]
])
for box, fn in boxes:
box = np.round(size_array * box).astype(np.int)
print(box)
if self.type == 'image':
image_box = np.ascontiguousarray(
image[box[0]:box[2], box[1]:box[3], :])
yield fn, imfeat.image_tostring(image_box, 'png')
elif self.type == 'feature':
image_box = np.ascontiguousarray(
image[box[0]:box[2], box[1]:box[3], :])
yield fn, feature.compute_patch(image_box)
elif self.type == 'box':
image2 = image.copy()
cv2.rectangle(image2, (box[1], box[0]),
(box[3], box[2]), (0, 255, 0), 4)
yield fn, imfeat.image_tostring(image2, 'jpg')
else:
raise ValueError(self.type)
def _action_handle(function, params, image):
print('Action[%s]' % function)
try:
ff = FEATURE_FUN[function]
image = imfeat.resize_image_max_side(image, 320) # TODO: Expose this
return {'feature': ff(**params)(image).tolist()}
except KeyError:
pass
try:
sf = SEARCH_FUN[function]
image = imfeat.resize_image_max_side(image, 320) # TODO: Expose this
print(sf.feature)
out = {'results': sf.analyze_cropped(image)}
if function == 'see/search/masks':
out['classes'] = CLASS_COLORS
return out
except KeyError:
pass
try:
cf = CLASSIFY_FUN[function]
image = imfeat.resize_image_max_side(image, 320) # TODO: Expose this
return {'results': cf(image)}
except KeyError:
pass
if function == 'see/texton' or function == 'see/texton_ilp':
image = cv2.resize(image, (320, int(image.shape[0] * 320. / image.shape[1])))
image = np.ascontiguousarray(image)
semantic_masks = TEXTON(image)
if function == 'see/texton_ilp':
ilp_pred = CLASSIFY_FUN['see/classify/indoor'](imfeat.resize_image_max_side(image, 320))
try:
bin_index = [x for x, y in enumerate(ILP_WEIGHTS['bins']) if y >= ilp_pred][0]
except IndexError:
bin_index = ILP_WEIGHTS['ilp_tables'].shape[1]
if bin_index != 0:
bin_index -= 1
ilp_weights = ILP_WEIGHTS['ilp_tables'][:, bin_index]
print('ILP Pred[%s] Weights[%s]' % (ilp_pred, ilp_weights))
semantic_masks *= ilp_weights
#min_probability = float(params.get('min_probability', 0.5))
#semantic_masks = np.dstack([semantic_masks, np.ones_like(semantic_masks[:, :, 0]) * min_probability])
texton_argmax2 = np.argmax(semantic_masks, 2)
image_string = imfeat.image_tostring(COLORS_BGR[texton_argmax2], 'png')
out = {'argmax_pngb64': base64.b64encode(image_string)}
out['classes'] = CLASS_COLORS
return out
if function == 'see/colors':
image = cv2.resize(image, (320, int(image.shape[0] * 320. / image.shape[1])))
image = np.ascontiguousarray(image)
masks = COLOR_NAMING.make_feature_mask(image)
mask_argmax = np.argmax(masks, 2)
image_string = imfeat.image_tostring(COLOR_NAMING.color_values[mask_argmax], 'png')
return {'argmax_pngb64': base64.b64encode(image_string)}
if function == 'see/faces':
results = [map(float, x) for x in FACES._detect_faces(image)]
return {'faces': [{'tl_x': x[0], 'tl_y': x[1], 'width': x[2], 'height': x[3]}
for x in results]}
return {}
def test_tostring(self):
for fn in ['lena.jpg', 'lena.pgm', 'lena.ppm']:
for i in load_images(fn):
for ext in ['jpeg', 'png']:
o = imfeat.image_tostring(i, ext)
o2 = imfeat.image_tostring(imfeat.image_fromstring(imfeat.image_tostring(i, ext)), ext)
if ext == 'png':
np.testing.assert_equal(o, o2)
s = imfeat.image_fromstring(o)
s2 = imfeat.image_fromstring(o2)
# No more than 9% of the pixels differ by more than 3
self.assertLess(np.mean(np.abs(s - s2) > 3), .09)
def test_tostring(self):
for fn in ['lena.jpg', 'lena.pgm', 'lena.ppm']:
for i in load_images(fn):
for ext in ['jpeg', 'png']:
o = imfeat.image_tostring(i, ext)
o2 = imfeat.image_tostring(imfeat.image_fromstring(imfeat.image_tostring(i, ext)), ext)
if ext == 'png':
np.testing.assert_equal(o, o2)
s = imfeat.image_fromstring(o)
s2 = imfeat.image_fromstring(o2)
# No more than 9% of the pixels differ by more than 3
self.assertLess(np.mean(np.abs(s - s2) > 3), .09)
def main():
exemplars = sorted(pickle.load(open('exemplars.pkl')), key=lambda x: x[0][2], reverse=True)[:100]
with open('exemplars_best.pkl', 'w') as fp:
pickle.dump(exemplars, fp, -1)
hdfs_output = 'exemplarbank/output/%s/' % '1341790878.92'
#hadoopy.launch_frozen('/user/brandyn/aladdin_results/keyframe/9/keyframe', hdfs_output + 'frame_pred', 'predict_video_frame.py', cmdenvs=['EXEMPLARS=exemplars_best.pkl', 'CELL_SKIP=1'], remove_output=True, files=['exemplars_best.pkl'])
local_out = 'frame_preds/'
try:
shutil.rmtree(local_out)
except OSError:
pass
os.makedirs(local_out)
for num, (data, (pyramid, num_boxes)) in enumerate(hadoopy.readtb(hdfs_output + 'frame_pred')):
if np.sum(pyramid):
pyramid_norm = pyramid / float(num_boxes)
pyramid_prob = np.sqrt(pyramid / float(np.max(pyramid)))
p = np.sum(pyramid_norm)
f = imfeat.image_fromstring(data['frame'])
pyramid_prob_frame = cv2.resize(pyramid_prob, (f.shape[1], f.shape[0]))
pyramid_prob_frame_color = COLORS[(pyramid_prob_frame * 255).astype(np.int), :]
alpha = .5
beta = alpha * pyramid_prob_frame
beta = beta.reshape((beta.shape[0], beta.shape[1], 1))
else:
beta = 0.
f = ((1 - beta) * f + beta * pyramid_prob_frame_color).astype(np.uint8)
print(p)
open(local_out + '%f-%d.jpg' % (p, num), 'w').write(imfeat.image_tostring(f, 'jpg'))
def map(self, key, value):
"""
Args:
key: Image name
value: Image as jpeg byte data
Yields:
A tuple in the form of (key, value)
key: (Image name, (x, y, w, h))
value: face image (.png)
"""
try:
image = imfeat.image_fromstring(value, {
'type': 'numpy',
'dtype': 'uint8',
'mode': 'gray'
})
image_color = imfeat.image_fromstring(value, {
'type': 'numpy',
'dtype': 'uint8',
'mode': 'bgr'
})
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
return
faces = _detect_faces(image, self._cascade)
for x, y, w, h in faces:
yield (key, (x, y, w, h)), imfeat.image_tostring(
image_color[y:y + h, x:x + w, :], '.png')
def image():
print_request()
# TODO(Cleanup)
image, params = _get_image()
image = imfeat.resize_image_max_side(image, 320)
image_string = imfeat.image_tostring(image, 'jpg')
return {'jpgb64': base64.b64encode(image_string)}
def image():
print_request()
# TODO(Cleanup)
image, params = _get_image()
image = imfeat.resize_image_max_side(image, 320)
image_string = imfeat.image_tostring(image, 'jpg')
return {'jpgb64': base64.b64encode(image_string)}
def map(self, event_filename, video_data):
"""
Args:
event_filename: Tuple of (event, filename)
video_data: Binary video data
Yields:
A tuple in the form of ((event, filename), value) where value is a dict
with contents
prev_frame_time:
prev_frame_num:
prev_frame:
frame_time:
frame_num:
frame:
"""
ext = '.' + event_filename[1].rsplit('.')[1]
with tempfile.NamedTemporaryFile(suffix=ext) as fp:
fp.write(video_data)
fp.flush()
prev_frame_time = None
prev_frame_num = None
prev_frame = None
try:
for (frame_num, frame_time, frame), iskeyframe in self.kf(
viderator.frame_iter(fp.name,
frame_skip=self.frame_skip,
frozen=True)):
if self.max_time < frame_time:
break
if iskeyframe and prev_frame is not None:
yield event_filename, {
'prev_frame_time': prev_frame_time,
'prev_frame_num': prev_frame_num,
'prev_frame':
imfeat.image_tostring(prev_frame, 'JPEG'),
'frame_time': frame_time,
'frame_num': frame_num,
'frame': imfeat.image_tostring(frame, 'JPEG')
}
prev_frame_num = frame_num
prev_frame = frame
except Exception, e:
print(e)
hadoopy.counter('VIDEO_ERROR', 'FFMPEGCantParse')
def map(self, image_id, image_binary):
try:
boxes = self.image_box_fns[image_id]
except KeyError:
pass
else:
image = imfeat.image_fromstring(image_binary)
size_array = np.array([image.shape[0], image.shape[1], image.shape[0], image.shape[1]])
scale_boxes = {}
for box, fn in boxes:
fbox = size_array * box
scale = int(np.round(np.log2((fbox[2] - fbox[0]) / feature.PATCH_SIZE)))
scale_check = int(np.round(np.log2((fbox[3] - fbox[1]) / feature.PATCH_SIZE)))
if scale != scale_check:
raise ValueError("Box is not square.")
scale_boxes.setdefault(scale, []).append((box, fn))
# Order boxes and fn's by scale
for scale in range(max(scale_boxes.keys()) + 1):
if scale > 0:
height, width = np.array(image.shape[:2]) / 2
image = image[: height * 2, : width * 2, :]
if min(width, height) < 1:
raise ValueError("Image is too small")
image = cv2.resize(image, (width, height))
if image is None: # NOTE(brandyn): It is too small
raise ValueError("Image is too small")
try:
boxes = scale_boxes[scale]
except KeyError:
continue
size_array = np.array([image.shape[0], image.shape[1], image.shape[0], image.shape[1]])
for box, fn in boxes:
box = np.round(size_array * box).astype(np.int)
print(box)
if self.type == "image":
image_box = np.ascontiguousarray(image[box[0] : box[2], box[1] : box[3], :])
yield fn, imfeat.image_tostring(image_box, "png")
elif self.type == "feature":
image_box = np.ascontiguousarray(image[box[0] : box[2], box[1] : box[3], :])
yield fn, feature.compute_patch(image_box)
elif self.type == "box":
image2 = image.copy()
cv2.rectangle(image2, (box[1], box[0]), (box[3], box[2]), (0, 255, 0), 4)
yield fn, imfeat.image_tostring(image2, "jpg")
else:
raise ValueError(self.type)
def reduce(self, key, values):
"""
For this to work we need a modified partitioner on the substring before the first tab.
This method operates on a single tile at level 0, and more tiles at higher zoom levels (4x each level).
Args:
key: (tile_id, subtile_id)
values: Iterator of [dist, images] where images are power of 2 JPEG
images in descending order by size
Yields:
Tuple of (key, value) where
key: Tile name
value: JPEG Image Data
"""
# Select minimum distance image, throw away score, and order images from smallest to largest powers of 2
self._sub_tiles[key] = min(values, key=lambda x: x[0])[1][::-1]
# As the images were JPEG, we need to make them arrays again
self._sub_tiles[key] = [
imfeat.image_fromstring(x) for x in self._sub_tiles[key]
]
# If we don't have all of the necessary subtiles.
if len(self._sub_tiles) != _subtiles_per_tile:
return
self._verify_subtile_keys()
for level in range(_levels):
# Each image is smaller than the tile
scale = 2**level
num_tiles = scale * scale
subtiles_per_tile_len = _subtiles_per_tile_length / scale
subtiles_per_tile = subtiles_per_tile_len**2
subtile_len = _subtile_length * scale
cur_subtiles = [(self._find_output(key, scale,
subtiles_per_tile_len,
subtile_len), images[level])
for key, images in self._sub_tiles.items()]
cur_subtiles.sort(key=lambda x: x[0])
cur_tile = np.zeros((_tile_length, _tile_length, 3),
dtype=np.uint8)
assert len(cur_subtiles) / subtiles_per_tile == num_tiles
cur_outtile = None
for subtile_ind, ((xouttile, youttile, xoffset, yoffset),
image) in enumerate(cur_subtiles):
if cur_outtile is None:
cur_outtile = (xouttile, youttile)
assert cur_outtile == (xouttile, youttile)
#cur_tile.paste(image, (xoffset, yoffset)) # TODO Suspect
cur_tile[yoffset:yoffset + image.shape[0],
xoffset:xoffset + image.shape[1], :] = image
if not (subtile_ind + 1) % subtiles_per_tile:
tile_name = '%d_%d_%d.jpg' % (level, xouttile, youttile)
yield tile_name, imfeat.image_tostring(cur_tile, 'jpg')
cur_tile = np.zeros((_tile_length, _tile_length, 3),
dtype=np.uint8)
cur_outtile = None
self._sub_tiles = {}
def map(self, event_filename, video_data):
"""
Args:
event_filename: Tuple of (event, filename)
video_data: Binary video data
Yields:
A tuple in the form of ((event, filename), value) where value is a dict
with contents
prev_frame_time:
prev_frame_num:
prev_frame:
frame_time:
frame_num:
frame:
"""
ext = '.' + event_filename[1].rsplit('.')[1]
with tempfile.NamedTemporaryFile(suffix=ext) as fp:
fp.write(video_data)
fp.flush()
prev_frame_time = None
prev_frame_num = None
prev_frame = None
try:
for (frame_num, frame_time, frame), iskeyframe in self.kf(viderator.frame_iter(fp.name,
frame_skip=self.frame_skip,
frozen=True)):
if self.max_time < frame_time:
break
if iskeyframe and prev_frame is not None:
yield event_filename, {'prev_frame_time': prev_frame_time,
'prev_frame_num': prev_frame_num,
'prev_frame': imfeat.image_tostring(prev_frame, 'JPEG'),
'frame_time': frame_time,
'frame_num': frame_num,
'frame': imfeat.image_tostring(frame, 'JPEG')}
prev_frame_num = frame_num
prev_frame = frame
except Exception, e:
print(e)
hadoopy.counter('VIDEO_ERROR', 'FFMPEGCantParse')
def load_data_iter(local_inputs):
# Push labeled samples to HDFS
unique_ids = set()
for fn in local_inputs:
try:
image_data = imfeat.image_tostring(cleanup_image(cv2.imread(fn)), '.jpg')
except (ValueError, IndexError):
continue
image_id = hashlib.md5(image_data).hexdigest()
if image_id not in unique_ids:
unique_ids.add(image_id)
yield image_id, image_data
def get_content(content_id):
if content_id.endswith('.b16.html'):
return '<img src="/content/%s.jpg" />' % base64.b16decode(content_id[:-9])
image_data = open(base64.b16decode(content_id[:content_id.find('.b16')])).read()
if content_id.endswith('.b16.thumb.jpg'):
try:
return CACHE[content_id]
except KeyError:
out_data = imfeat.image_tostring(imfeat.resize_image(imfeat.image_fromstring(image_data), 200, 200), 'JPEG')
CACHE[content_id] = out_data
return out_data
return image_data
def close(self):
"""
Yields:
Tuple of (key, value) where
key: tile_id\tsubtile_id (easily parsable by the
KeyFieldBasedPartitioner)
value: [dist, images] where images are power of 2 JPEG images
in descending order by size
"""
assert sorted(self.target_tiles.keys()) == sorted(self.min_dists.keys())
for key, (dist, images) in self.min_dists.items(): # sorted(
images_jpg = [imfeat.image_tostring(x, 'jpg') for x in images] # JPEG's are much smaller
yield key, [dist, images_jpg]
def load_data_iter(local_inputs):
# Push labeled samples to HDFS
unique_ids = set()
for fn in local_inputs:
try:
image_data = imfeat.image_tostring(cleanup_image(cv2.imread(fn)),
'.jpg')
except (ValueError, IndexError):
continue
image_id = hashlib.md5(image_data).hexdigest()
if image_id not in unique_ids:
unique_ids.add(image_id)
yield image_id, image_data
def reduce(self, key, values):
"""
For this to work we need a modified partitioner on the substring before the first tab.
This method operates on a single tile at level 0, and more tiles at higher zoom levels (4x each level).
Args:
key: (tile_id, subtile_id)
values: Iterator of [dist, images] where images are power of 2 JPEG
images in descending order by size
Yields:
Tuple of (key, value) where
key: Tile name
value: JPEG Image Data
"""
# Select minimum distance image, throw away score, and order images from smallest to largest powers of 2
self._sub_tiles[key] = min(values, key=lambda x: x[0])[1][::-1]
# As the images were JPEG, we need to make them arrays again
self._sub_tiles[key] = [imfeat.image_fromstring(x) for x in self._sub_tiles[key]]
# If we don't have all of the necessary subtiles.
if len(self._sub_tiles) != _subtiles_per_tile:
return
self._verify_subtile_keys()
for level in range(_levels):
# Each image is smaller than the tile
scale = 2 ** level
num_tiles = scale * scale
subtiles_per_tile_len = _subtiles_per_tile_length / scale
subtiles_per_tile = subtiles_per_tile_len ** 2
subtile_len = _subtile_length * scale
cur_subtiles = [(self._find_output(key, scale, subtiles_per_tile_len, subtile_len), images[level])
for key, images in self._sub_tiles.items()]
cur_subtiles.sort(key=lambda x: x[0])
cur_tile = np.zeros((_tile_length, _tile_length, 3), dtype=np.uint8)
assert len(cur_subtiles) / subtiles_per_tile == num_tiles
cur_outtile = None
for subtile_ind, ((xouttile, youttile, xoffset, yoffset), image) in enumerate(cur_subtiles):
if cur_outtile is None:
cur_outtile = (xouttile, youttile)
assert cur_outtile == (xouttile, youttile)
#cur_tile.paste(image, (xoffset, yoffset)) # TODO Suspect
cur_tile[yoffset:yoffset + image.shape[0], xoffset:xoffset + image.shape[1], :] = image
if not (subtile_ind + 1) % subtiles_per_tile:
tile_name = '%d_%d_%d.jpg' % (level, xouttile, youttile)
yield tile_name, imfeat.image_tostring(cur_tile, 'jpg')
cur_tile = np.zeros((_tile_length, _tile_length, 3), dtype=np.uint8)
cur_outtile = None
self._sub_tiles = {}
def close(self):
"""
Yields:
Tuple of (key, value) where
key: tile_id\tsubtile_id (easily parsable by the
KeyFieldBasedPartitioner)
value: [dist, images] where images are power of 2 JPEG images
in descending order by size
"""
assert sorted(self.target_tiles.keys()) == sorted(
self.min_dists.keys())
for key, (dist, images) in self.min_dists.items(): # sorted(
images_jpg = [imfeat.image_tostring(x, 'jpg')
for x in images] # JPEG's are much smaller
yield key, [dist, images_jpg]
def map(self, name, image_data):
try:
image = imfeat.image_fromstring(image_data)
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
return
if self.filter_side is not None and min(image.shape[0], image.shape[1]) < self.filter_side:
hadoopy.counter('DATA_ERRORS', 'ImageTooSmallPre')
return
if self.max_side is not None:
image = imfeat.resize_image_max_side(image, self.max_side)
if self.filter_side is not None and min(image.shape[0], image.shape[1]) < self.filter_side:
hadoopy.counter('DATA_ERRORS', 'ImageTooSmallPost')
return
yield name, imfeat.image_tostring(image, 'jpg')
def map(self, name, image_data):
try:
image = imfeat.image_fromstring(image_data)
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
return
if self.filter_side is not None and min(
image.shape[0], image.shape[1]) < self.filter_side:
hadoopy.counter('DATA_ERRORS', 'ImageTooSmallPre')
return
if self.max_side is not None:
image = imfeat.resize_image_max_side(image, self.max_side)
if self.filter_side is not None and min(
image.shape[0], image.shape[1]) < self.filter_side:
hadoopy.counter('DATA_ERRORS', 'ImageTooSmallPost')
return
yield name, imfeat.image_tostring(image, 'jpg')
def map(self, key, value):
"""
Args:
key: Image name
value: Image as jpeg byte data
Yields:
A tuple in the form of (key, value)
key: (Image name, (x, y, w, h))
value: face image (.png)
"""
try:
image = imfeat.image_fromstring(value, {'type': 'numpy', 'dtype': 'uint8', 'mode': 'gray'})
image_color = imfeat.image_fromstring(value, {'type': 'numpy', 'dtype': 'uint8', 'mode': 'bgr'})
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
return
faces = _detect_faces(image, self._cascade)
for x, y, w, h in faces:
yield (key, (x, y, w, h)), imfeat.image_tostring(image_color[y:y + h, x:x + w, :], '.png')
def map(self, event_filename, video_data):
"""
Args:
event_filename: Tuple of (event, filename)
video_data: Binary video data
Yields:
A tuple in the form of ((event, filename, frame_num, frame_time), frame_data)
"""
ext = '.' + event_filename[1].rsplit('.')[1]
event, filename = event_filename
out_count = 0
with tempfile.NamedTemporaryFile(suffix=ext) as fp:
fp.write(video_data)
fp.flush()
try:
for frame_num, frame_time, frame in viderator.frame_iter(
fp.name, frozen=True, frame_skip=self.frame_skip):
if frame_num >= self.max_frames_per_video:
break
frame_orig = frame
if self.remove_bars:
sz = self.remove_bars.find_bars(frame)
frame = frame[sz[0]:sz[1], sz[2]:sz[3], :]
if not frame.size: # Empty
continue
if self._feat.predict(frame):
if self.output_frame:
yield (event, filename, frame_num,
frame_time), imfeat.image_tostring(
frame_orig, 'JPEG')
else:
yield (event, filename, frame_num, frame_time), ''
out_count += 1
if out_count >= self.max_outputs_per_video:
break
except IOError:
hadoopy.counter('PICARUS', 'CantProcessVideo')
def map(self, event_filename, video_data):
"""
Args:
event_filename: Tuple of (event, filename)
video_data: Binary video data
Yields:
A tuple in the form of ((event, filename, frame_num, frame_time), frame_data)
"""
ext = '.' + event_filename[1].rsplit('.')[1]
event, filename = event_filename
out_count = 0
with tempfile.NamedTemporaryFile(suffix=ext) as fp:
fp.write(video_data)
fp.flush()
try:
for frame_num, frame_time, frame in viderator.frame_iter(fp.name,
frozen=True,
frame_skip=self.frame_skip):
if frame_num >= self.max_frames_per_video:
break
frame_orig = frame
if self.remove_bars:
sz = self.remove_bars.find_bars(frame)
frame = frame[sz[0]:sz[1], sz[2]:sz[3], :]
if not frame.size: # Empty
continue
if self._feat.predict(frame):
if self.output_frame:
yield (event, filename, frame_num, frame_time), imfeat.image_tostring(frame_orig, 'JPEG')
else:
yield (event, filename, frame_num, frame_time), ''
out_count += 1
if out_count >= self.max_outputs_per_video:
break
except IOError:
hadoopy.counter('PICARUS', 'CantProcessVideo')
def map(self, event_filename, video_data):
"""
Args:
event_filename: Tuple of (event, filename)
video_data: Binary video data
Yields:
A tuple in the form of ((event, filename, frame_num, frame_time), frame_data)
"""
ext = '.' + event_filename[1].rsplit('.')[1]
event, filename = event_filename
heap = [(float('-inf'), None)] * self.max_outputs
with tempfile.NamedTemporaryFile(suffix=ext) as fp:
fp.write(video_data)
fp.flush()
sys.stderr.write('Prevideo\n')
try:
for frame_num, frame_time, frame in viderator.frame_iter(fp.name,
frozen=True,
frame_skip=self.frame_skip):
sys.stderr.write('FrameNum[%d]\n' % frame_num)
if frame_num >= self.max_frames_per_video:
break
frame_orig = frame
if self.remove_bars:
sz = self.remove_bars.find_bars(frame)
frame = frame[sz[0]:sz[1], sz[2]:sz[3], :]
if not frame.size: # Empty
continue
st = time.time()
c = self._feat(frame)[0]
sys.stderr.write('FrameTime[%f]\n' % (time.time() - st))
print('FrameTime[%f]' % (time.time() - st))
if c > heap[0][0]:
if self.output_frame:
heapq.heappushpop(heap,
(c, ((event, filename, frame_num, frame_time), imfeat.image_tostring(frame_orig, 'JPEG'))))
else:
heapq.heappushpop(heap,
(c, ((event, filename, frame_num, frame_time), '')))
except IOError:
hadoopy.counter('PICARUS', 'CantProcessVideo')
for x in heap[-self.max_outputs_per_video:]:
heapq.heappushpop(self.heap, x)
def _map(self, row, image_binary):
try:
image = imfeat.image_fromstring(image_binary)
yield row, imfeat.image_tostring(imfeat.resize_image_max_side(image, self.max_side), 'jpg')
except:
hadoopy.counter('DATA_ERRORS', 'ImageLoadError')
def _action_handle(function, params, image):
print('Action[%s]' % function)
try:
ff = FEATURE_FUN[function]
image = imfeat.resize_image_max_side(image, 320) # TODO: Expose this
return {'feature': ff(**params)(image).tolist()}
except KeyError:
pass
try:
sf = SEARCH_FUN[function]
image = imfeat.resize_image_max_side(image, 320) # TODO: Expose this
print(sf.feature)
out = {'results': sf.analyze_cropped(image)}
if function == 'see/search/masks':
out['classes'] = CLASS_COLORS
return out
except KeyError:
pass
try:
cf = CLASSIFY_FUN[function]
image = imfeat.resize_image_max_side(image, 320) # TODO: Expose this
return {'results': cf(image)}
except KeyError:
pass
if function == 'see/texton' or function == 'see/texton_ilp':
image = cv2.resize(image,
(320, int(image.shape[0] * 320. / image.shape[1])))
image = np.ascontiguousarray(image)
semantic_masks = TEXTON(image)
if function == 'see/texton_ilp':
ilp_pred = CLASSIFY_FUN['see/classify/indoor'](
imfeat.resize_image_max_side(image, 320))
try:
bin_index = [
x for x, y in enumerate(ILP_WEIGHTS['bins'])
if y >= ilp_pred
][0]
except IndexError:
bin_index = ILP_WEIGHTS['ilp_tables'].shape[1]
if bin_index != 0:
bin_index -= 1
ilp_weights = ILP_WEIGHTS['ilp_tables'][:, bin_index]
print('ILP Pred[%s] Weights[%s]' % (ilp_pred, ilp_weights))
semantic_masks *= ilp_weights
#min_probability = float(params.get('min_probability', 0.5))
#semantic_masks = np.dstack([semantic_masks, np.ones_like(semantic_masks[:, :, 0]) * min_probability])
texton_argmax2 = np.argmax(semantic_masks, 2)
image_string = imfeat.image_tostring(COLORS_BGR[texton_argmax2], 'png')
out = {'argmax_pngb64': base64.b64encode(image_string)}
out['classes'] = CLASS_COLORS
return out
if function == 'see/colors':
image = cv2.resize(image,
(320, int(image.shape[0] * 320. / image.shape[1])))
image = np.ascontiguousarray(image)
masks = COLOR_NAMING.make_feature_mask(image)
mask_argmax = np.argmax(masks, 2)
image_string = imfeat.image_tostring(
COLOR_NAMING.color_values[mask_argmax], 'png')
return {'argmax_pngb64': base64.b64encode(image_string)}
if function == 'see/faces':
results = [map(float, x) for x in FACES._detect_faces(image)]
return {
'faces': [{
'tl_x': x[0],
'tl_y': x[1],
'width': x[2],
'height': x[3]
} for x in results]
}
return {}