def _get_video_feats(self, segment_dict):
start, end = segment_dict['start'], segment_dict['end']
video_name = segment_dict['video-name']
i3d_feat = breakfast.read_i3d_data(video_name,
window=[start, end],
i3d_length=self.i3d_length)
idx = np.random.choice(np.arange(len(i3d_feat)))
i3d_feat = torch.from_numpy(i3d_feat[idx])
return i3d_feat
def __getitem__(self, idx):
segment = self.segments[idx]
video_name = segment['video-name']
start, end = segment['start'], segment['end']
i3d_feat = breakfast.read_i3d_data(video_name,
window=[start, end],
i3d_length=self.i3d_length)
sample_idxs = self._get_sample_idxs(start, end)
selected = i3d_feat[sample_idxs]
selected = torch.from_numpy(selected)
return selected
def __getitem__(self, idx):
segment = self.segments[idx]
video_name = segment['video-name']
start, end = segment['start'], segment['end']
i3d_feat = breakfast.read_i3d_data(video_name,
window=[start, end],
i3d_length=self.i3d_length)
i3d_feat = i3d_feat[::self.stride]
i3d_feat = torch.from_numpy(i3d_feat)
segment_len = i3d_feat.shape[0]
return i3d_feat, segment_len
def __getitem__(self, idx):
segment_dict = self.segments[idx]
logit = self.segment_logits[idx]
video_name = segment_dict['video-name']
start, end = segment_dict['start'], segment_dict['end']
assert start < end, '{0} has errors, logit {1}'.format(
video_name, logit)
i3d_feat = breakfast.read_i3d_data(video_name,
window=[start, end],
i3d_length=self.i3d_length)
assert len(i3d_feat) > 0, '{0} has length {1}, logit {2}'.format(
video_name, len(i3d_feat), logit)
selected_idx = np.random.choice(np.arange(len(i3d_feat)))
selected = torch.from_numpy(i3d_feat[selected_idx])
return selected, logit
def _parse_split_data(split, feat_len):
segments, labels, logits = breakfast.get_data(split)
valid_segments = []
valid_labels = []
valid_logits = []
for i, segment in enumerate(tqdm(segments)):
start, end = segment['start'], segment['end']
i3d_feats = breakfast.read_i3d_data(segment['video-name'],
window=[start, end],
i3d_length=feat_len)
if len(i3d_feats
) > 0 and 48 > logits[i] > 0: # remove walk in and walk out.
segment['end'] = segment['start'] + len(i3d_feats)
valid_segments.append(segment)
valid_labels.append(labels[i])
valid_logits.append(logits[i])
return [valid_segments, valid_labels, valid_logits]
def __getitem__(self, idx):
segment_dict = self.segments[idx]
logit = self.segment_logits[idx]
video_name = segment_dict['video-name']
start, end = segment_dict['start'], segment_dict['end']
assert start < end, '{0} has errors, logit {1}'.format(
video_name, logit)
i3d_feat = breakfast.read_i3d_data(video_name,
window=[start, end],
i3d_length=self.i3d_length)
assert len(i3d_feat) > 0, '{0} has length {1}, logit {2}'.format(
video_name, len(i3d_feat), logit)
i3d_feat = i3d_feat[::self.stride]
i3d_feat = torch.from_numpy(i3d_feat)
n_feats = i3d_feat.shape[0]
return i3d_feat, n_feats, logit
