def to_aggregated_dataset(self, aggregate_scores, **kwargs):
newone = empty_object()
# systematically copy fields, e.g. dataset_name, yuv_fmt, width, height, ...
for key in self.dataset.__dict__.keys():
if not key.startswith(
'__'): # filter out those e.g. __builtin__ ...
setattr(newone, key, getattr(self.dataset, key))
if 'quality_width' in kwargs and kwargs['quality_width'] is not None:
newone.quality_width = kwargs['quality_width']
elif hasattr(self.dataset, 'quality_width'):
newone.quality_width = self.dataset.quality_width
if 'quality_height' in kwargs and kwargs['quality_height'] is not None:
newone.quality_height = kwargs['quality_height']
elif hasattr(self.dataset, 'quality_height'):
newone.quality_height = self.dataset.quality_height
if 'resampling_type' in kwargs and kwargs[
'resampling_type'] is not None:
newone.resampling_type = kwargs['resampling_type']
elif hasattr(self.dataset, 'resampling_type'):
newone.resampling_type = self.dataset.resampling_type
# ref_videos: deepcopy
newone.ref_videos = copy.deepcopy(self.dataset.ref_videos)
# dis_videos: use input aggregate scores
dis_videos = []
assert len(self.dataset.dis_videos) == len(aggregate_scores)
for dis_video, score in zip(self.dataset.dis_videos, aggregate_scores):
dis_video2 = copy.deepcopy(dis_video)
if 'os' in dis_video2: # remove 'os' - opinion score
del dis_video2['os']
dis_video2['groundtruth'] = score
dis_videos.append(dis_video2)
# add scores std if available
if 'scores_std' in kwargs and kwargs['scores_std'] is not None:
assert len(dis_videos) == len(kwargs['scores_std'])
for dis_video, score_std in zip(dis_videos, kwargs['scores_std']):
dis_video['groundtruth_std'] = score_std
if 'aggregate_content_ids' in kwargs and kwargs[
'aggregate_content_ids'] is not None:
dis_videos = filter(
lambda dis_video: dis_video['content_id'] in kwargs[
'aggregate_content_ids'], dis_videos)
if 'aggregate_asset_ids' in kwargs and kwargs[
'aggregate_asset_ids'] is not None:
dis_videos = filter(
lambda dis_video: dis_video['asset_id'] in kwargs[
'aggregate_asset_ids'], dis_videos)
newone.dis_videos = dis_videos
return newone
def to_persubject_dataset(self, quality_scores, **kwargs):
import math
newone = empty_object()
# systematically copy fields, e.g. dataset_name, yuv_fmt, width, height, ...
for key in self.dataset.__dict__.keys():
if not key.startswith(
'__'): # filter out those e.g. __builtin__ ...
setattr(newone, key, getattr(self.dataset, key))
if 'quality_width' in kwargs and kwargs['quality_width'] is not None:
newone.quality_width = kwargs['quality_width']
elif hasattr(self.dataset, 'quality_width'):
newone.quality_width = self.dataset.quality_width
if 'quality_height' in kwargs and kwargs['quality_height'] is not None:
newone.quality_height = kwargs['quality_height']
elif hasattr(self.dataset, 'quality_height'):
newone.quality_height = self.dataset.quality_height
if 'resampling_type' in kwargs and kwargs[
'resampling_type'] is not None:
newone.resampling_type = kwargs['resampling_type']
elif hasattr(self.dataset, 'resampling_type'):
newone.resampling_type = self.dataset.resampling_type
# ref_videos: deepcopy
newone.ref_videos = copy.deepcopy(self.dataset.ref_videos)
# dis_videos: use input aggregate scores
dis_videos = []
for dis_video, quality_score in zip(self.dataset.dis_videos,
quality_scores):
assert 'os' in dis_video
# quality_score should be a 1-D array with (processed) per-subject scores
assert hasattr(quality_score, '__len__')
# new style: opinion is specified as a dict: user -> score. In this
# case, quality_score may contain nan. In this case: filter them out
if isinstance(dis_video['os'], dict):
quality_score = filter(lambda x: not math.isnan(x),
quality_score)
assert len(dis_video['os']) == len(quality_score)
for persubject_score in quality_score:
dis_video2 = copy.deepcopy(dis_video)
if 'os' in dis_video2: # remove 'os' - opinion score
del dis_video2['os']
dis_video2['groundtruth'] = persubject_score
dis_videos.append(dis_video2)
newone.dis_videos = dis_videos
return newone
def _prepare_new_dataset(self, kwargs):
newone = empty_object()
# systematically copy fields, e.g. dataset_name, yuv_fmt, width, height, ...
for key in self.dataset.__dict__.keys():
if not key.startswith('__'): # filter out those e.g. __builtin__ ...
setattr(newone, key, getattr(self.dataset, key))
if 'quality_width' in kwargs and kwargs['quality_width'] is not None:
newone.quality_width = kwargs['quality_width']
elif hasattr(self.dataset, 'quality_width'):
newone.quality_width = self.dataset.quality_width
if 'quality_height' in kwargs and kwargs['quality_height'] is not None:
newone.quality_height = kwargs['quality_height']
elif hasattr(self.dataset, 'quality_height'):
newone.quality_height = self.dataset.quality_height
if 'resampling_type' in kwargs and kwargs['resampling_type'] is not None:
newone.resampling_type = kwargs['resampling_type']
elif hasattr(self.dataset, 'resampling_type'):
newone.resampling_type = self.dataset.resampling_type
return newone
def to_dataset(self):
"""
Override DatasetReader.to_dataset(). Need to overwrite dis_video['os']
"""
newone = empty_object()
newone.__dict__.update(self.dataset.__dict__)
# deep copy ref_videos and dis_videos
newone.ref_videos = copy.deepcopy(self.dataset.ref_videos)
newone.dis_videos = copy.deepcopy(self.dataset.dis_videos)
# overwrite dis_video['os']
score_mtx = self.opinion_score_2darray
num_videos, num_subjects = score_mtx.shape
assert num_videos == len(newone.dis_videos)
for scores, dis_video in zip(score_mtx, newone.dis_videos):
dis_video['os'] = list(scores)
return newone
