def make_dataset(root_path, annotation_path, subset, n_samples_for_each_video,
sample_duration):
data = load_annotation_data(annotation_path)
video_names, annotations = get_video_names_and_annotations(data, subset)
class_to_idx = get_class_labels(data)
idx_to_class = {}
for name, label in class_to_idx.items():
idx_to_class[label] = name
dataset = []
for i in range(len(video_names)):
# if i % 1000 == 0:
# print('dataset loading [{}/{}]'.format(i, len(video_names)))
video_path = os.path.join(root_path, video_names[i])
if not os.path.exists(video_path):
continue
n_frames_file_path = os.path.join(video_path, 'n_frames')
n_frames = int(load_value_file(n_frames_file_path))
if n_frames <= 0:
continue
begin_t = 1
end_t = n_frames
sample = {
'video': video_path,
'segment': [begin_t, end_t],
'n_frames': n_frames,
'video_id': video_names[i].split('/')[1]
}
if len(annotations) != 0:
sample['label'] = class_to_idx[annotations[i]['label']]
else:
sample['label'] = -1
if n_samples_for_each_video == 1:
sample['frame_indices'] = list(range(1, n_frames + 1))
dataset.append(sample)
else:
if n_samples_for_each_video > 1:
step = max(
1,
math.ceil((n_frames - 1 - sample_duration) /
(n_samples_for_each_video - 1)))
else:
step = sample_duration
for j in range(1, n_frames, step):
sample_j = copy.deepcopy(sample)
sample_j['frame_indices'] = list(
range(j, min(n_frames + 1, j + sample_duration)))
dataset.append(sample_j)
return dataset, idx_to_class
def make_untrimmed_dataset(root_path, annotation_path, subset,
n_samples_for_each_video, sample_duration):
data = load_annotation_data(annotation_path)
video_names, _ = get_video_names_and_annotations(data, subset)
class_to_idx = get_class_labels(data)
idx_to_class = {}
for name, label in class_to_idx.items():
idx_to_class[label] = name
dataset = []
for i in range(len(video_names)):
if i % 1000 == 0:
print('dataset loading [{}/{}]'.format(i, len(video_names)))
video_path = os.path.join(root_path, video_names[i])
if not os.path.exists(video_path):
continue
fps_file_path = os.path.join(video_path, 'fps')
fps = load_value_file(fps_file_path)
begin_t = 1
end_t = get_end_t(video_path)
n_frames = end_t - begin_t
sample = {
'video': video_path,
'segment': [begin_t, end_t],
'fps': fps,
'video_id': video_names[i][2:]
}
if n_samples_for_each_video >= 1:
step = max(
1,
math.ceil((n_frames - 1 - sample_duration) /
(n_samples_for_each_video - 1)))
else:
step = sample_duration
for j in range(begin_t, end_t, step):
sample_j = copy.deepcopy(sample)
frame_indices = list(range(j, j + sample_duration))
frame_indices = modify_frame_indices(sample_j['video'],
frame_indices)
if len(frame_indices) < 16:
continue
sample_j['frame_indices'] = frame_indices
dataset.append(sample_j)
return dataset, idx_to_class
def make_dataset_human_det(root_path, annotation_path, subset,
n_samples_for_each_video, sample_duration, dets):
data = load_annotation_data(annotation_path)
video_names, annotations = get_video_names_and_annotations(data, subset)
class_to_idx = get_class_labels(data)
idx_to_class = {}
for name, label in class_to_idx.items():
idx_to_class[label] = name
num_samples_wo_filtering = 0
dataset = []
for i in range(len(video_names)):
if i % 1000 == 0:
print('dataset loading [{}/{}]'.format(i, len(video_names)))
video_path = os.path.join(root_path, video_names[i])
if not os.path.exists(video_path):
continue
n_frames_file_path = os.path.join(video_path, 'n_frames')
n_frames = int(load_value_file(n_frames_file_path))
if n_frames <= 0:
continue
if n_samples_for_each_video == 1:
num_samples_wo_filtering += 1
else:
if n_samples_for_each_video > 1:
step = max(
1,
math.ceil((n_frames - 1 - sample_duration) /
(n_samples_for_each_video - 1)))
else:
step = sample_duration
num_samples_wo_filtering += np.arange(1, n_frames, step).shape[0]
cur_cls = video_path.split('/')[-2]
vid = video_path.split('/')[-1][:11]
if not is_video_valid_det(
dets[cur_cls][vid], n_frames, det_th=0.3, ratio_th=0.7):
continue
begin_t = 1
end_t = n_frames
sample = {
'video': video_path,
'segment': [begin_t, end_t],
'n_frames': n_frames,
'video_id': video_names[i][:-14].split('/')[1]
}
if len(annotations) != 0:
sample['label'] = class_to_idx[annotations[i]['label']]
else:
sample['label'] = -1
if n_samples_for_each_video == 1:
sample['frame_indices'] = list(range(1, n_frames + 1))
dataset.append(sample)
else:
if n_samples_for_each_video > 1:
step = max(
1,
math.ceil((n_frames - 1 - sample_duration) /
(n_samples_for_each_video - 1)))
else:
step = sample_duration
for j in range(1, n_frames, step):
sample_j = copy.deepcopy(sample)
sample_j['frame_indices'] = list(
range(j, min(n_frames + 1, j + sample_duration)))
dataset.append(sample_j)
print(
'len(dataset) after filtering the videos without sufficient detections: [{}/{}]'
.format(len(dataset), num_samples_wo_filtering))
# repeat the dataset so that the number of samples can be matched to the original dataset w/o filtering
if len(dataset) < num_samples_wo_filtering:
num_repeat = np.ceil(num_samples_wo_filtering / len(dataset)).astype(
np.int32)
dataset_repeat = copy.deepcopy(dataset)
for i in range(1, num_repeat):
dataset_shuffled = copy.deepcopy(dataset)
np.random.shuffle(dataset_shuffled)
dataset_repeat += dataset_shuffled
return dataset_repeat[:num_samples_wo_filtering], idx_to_class