def gen_detection_results(video_id, score_tp):
if len(score_tp[0].shape) == 3:
rel_prop = np.squeeze(score_tp[0], 0)
else:
rel_prop = score_tp[0]
# standardize regression scores
reg_scores = score_tp[3]
if reg_scores is None:
reg_scores = np.zeros((len(rel_prop), num_class, 2), dtype=np.float32)
reg_scores = reg_scores.reshape((-1, num_class, 2))
if top_k <= 0 and cls_score_dict is None:
combined_scores = softmax(score_tp[1])[:, 1:] * np.exp(score_tp[2])
for i in range(num_class):
loc_scores = reg_scores[:, i, 0][:, None]
dur_scores = reg_scores[:, i, 1][:, None]
try:
dataset_detections[i][video_id] = np.concatenate((
rel_prop, combined_scores[:, i][:, None], loc_scores, dur_scores), axis=1)
except:
print(i, rel_prop.shape, combined_scores.shape, reg_scores.shape)
raise
elif cls_score_dict is None:
combined_scores = softmax(score_tp[1][:, 1:]) * np.exp(score_tp[2])
keep_idx = np.argsort(combined_scores.ravel())[-top_k:]
for k in keep_idx:
cls = k % num_class
prop_idx = k // num_class
if video_id not in dataset_detections[cls]:
dataset_detections[cls][video_id] = np.array([
[rel_prop[prop_idx, 0], rel_prop[prop_idx, 1], combined_scores[prop_idx, cls],
reg_scores[prop_idx, cls, 0], reg_scores[prop_idx, cls, 1]]
])
else:
dataset_detections[cls][video_id] = np.vstack(
[dataset_detections[cls][video_id],
[rel_prop[prop_idx, 0], rel_prop[prop_idx, 1], combined_scores[prop_idx, cls],
reg_scores[prop_idx, cls, 0], reg_scores[prop_idx, cls, 1]]])
else:
if softmax_bf:
combined_scores = softmax(score_tp[1])[:, 1:] * np.exp(score_tp[2])
else:
combined_scores = score_tp[1][:, 1:] * np.exp(score_tp[2])
video_cls_score = cls_score_dict[os.path.splitext(os.path.basename(video_id))[0]]
for video_cls in np.argsort(video_cls_score,)[-args.cls_top_k:]:
loc_scores = reg_scores[:, video_cls, 0][:, None]
dur_scores = reg_scores[:, video_cls, 1][:, None]
try:
dataset_detections[video_cls][video_id] = np.concatenate((
rel_prop, combined_scores[:, video_cls][:, None], loc_scores, dur_scores), axis=1)
except:
print(video_cls, rel_prop.shape, combined_scores.shape, reg_scores.shape, loc_scores.shape, dur_scores.shape)
raise
def gen_detection_results(video_id, score_tp):
if len(score_tp[0].shape) == 3:
rel_prop = np.squeeze(score_tp[0], 0)
else:
rel_prop = score_tp[0]
# standardize regression scores
reg_scores = score_tp[3]
if reg_scores is None:
reg_scores = np.zeros((len(rel_prop), num_class, 2), dtype=np.float32)
reg_scores = reg_scores.reshape((-1, num_class, 2))
if top_k <= 0 and cls_score_dict is None:
combined_scores = softmax(score_tp[1])[:, 1:] * np.exp(score_tp[2])
for i in range(num_class):
loc_scores = reg_scores[:, i, 0][:, None]
dur_scores = reg_scores[:, i, 1][:, None]
try:
dataset_detections[i][video_id] = np.concatenate((
rel_prop, combined_scores[:, i][:, None], loc_scores, dur_scores), axis=1)
except:
print(i, rel_prop.shape, combined_scores.shape, reg_scores.shape)
raise
elif cls_score_dict is None:
combined_scores = softmax(score_tp[1][:, 1:]) * np.exp(score_tp[2])
keep_idx = np.argsort(combined_scores.ravel())[-top_k:]
for k in keep_idx:
cls = k % num_class
prop_idx = k // num_class
if video_id not in dataset_detections[cls]:
dataset_detections[cls][video_id] = np.array([
[rel_prop[prop_idx, 0], rel_prop[prop_idx, 1], combined_scores[prop_idx, cls],
reg_scores[prop_idx, cls, 0], reg_scores[prop_idx, cls, 1]]
])
else:
dataset_detections[cls][video_id] = np.vstack(
[dataset_detections[cls][video_id],
[rel_prop[prop_idx, 0], rel_prop[prop_idx, 1], combined_scores[prop_idx, cls],
reg_scores[prop_idx, cls, 0], reg_scores[prop_idx, cls, 1]]])
else:
if softmax_bf:
combined_scores = softmax(score_tp[1])[:, 1:] * np.exp(score_tp[2])
else:
combined_scores = score_tp[1][:, 1:] * np.exp(score_tp[2])
video_cls_score = cls_score_dict[os.path.splitext(os.path.basename(video_id))[0]]
for video_cls in np.argsort(video_cls_score,)[-args.cls_top_k:]:
loc_scores = reg_scores[:, video_cls, 0][:, None]
dur_scores = reg_scores[:, video_cls, 1][:, None]
try:
dataset_detections[video_cls][video_id] = np.concatenate((
rel_prop, combined_scores[:, video_cls][:, None], loc_scores, dur_scores), axis=1)
except:
print(video_cls, rel_prop.shape, combined_scores.shape, reg_scores.shape, loc_scores.shape, dur_scores.shape)
raise
def gen_detection_results(video_id, score_tp):
if len(score_tp[0].shape) == 3:
rel_prop = np.squeeze(score_tp[0], 0)
else:
rel_prop = score_tp[0]
# standardize regression scores
reg_scores = score_tp[3]
if reg_scores is None:
reg_scores = np.zeros((len(rel_prop), num_class, 2), dtype=np.float32)
reg_scores = reg_scores.reshape((-1, num_class, 2))
combined_scores = softmax(score_tp[1].squeeze()[:, :])
combined_scores = combined_scores[:, 1:]
combined_scores = combined_scores * np.exp(score_tp[2].squeeze())
keep_idx = np.argsort(combined_scores.ravel())[-top_k:]
for k in keep_idx:
cls = k % num_class
prop_idx = k // num_class
if video_id not in dataset_detections[cls]:
dataset_detections[cls][video_id] = np.array([[
rel_prop[prop_idx, 0], rel_prop[prop_idx, 1],
combined_scores[prop_idx, cls], reg_scores[prop_idx, cls, 0],
reg_scores[prop_idx, cls, 1]
]])
else:
dataset_detections[cls][video_id] = np.vstack([
dataset_detections[cls][video_id],
[
rel_prop[prop_idx, 0], rel_prop[prop_idx, 1],
combined_scores[prop_idx, cls],
reg_scores[prop_idx, cls, 0], reg_scores[prop_idx, cls, 1]
]
])
def gen_detection_results(video_id, score_tp):
if len(score_tp[0].shape) == 3:
rel_prop = np.squeeze(score_tp[0], 0)
else:
rel_prop = score_tp[0]
# standardize regression scores
reg_scores = score_tp[3]
if reg_scores is None:
reg_scores = np.zeros((len(rel_prop), num_class, 2), dtype=np.float32)
reg_scores = reg_scores.reshape((-1, num_class, 2))
if cls_score_dict is None:
combined_scores = softmax(score_tp[1][:, :])
combined_scores = combined_scores[:, 1:]
else:
combined_scores = softmax(cls_score_dict[video_id][1])[:, 1:]
if combined_scores.shape[1] < score_tp[2].shape[1]:
combined_scores = np.concatenate(
(combined_scores,
np.zeros([
len(combined_scores),
score_tp[2].shape[1] - combined_scores.shape[1]
])),
axis=1)
elif combined_scores.shape[1] > score_tp[2].shape[1]:
combined_scores = combined_scores[:, :score_tp[2].shape[1]]
if not args.no_comp:
combined_scores = combined_scores * np.exp(score_tp[2])
keep_idx = np.argsort(combined_scores.ravel())[-top_k:]
# pdb.set_trace()
delete_short = args.min_length is not None
if delete_short:
print('delete short proposals')
duration = dataset.video_dict[video_id].num_frames / 6
prop_duration = duration * (rel_prop[:, 1] - rel_prop[:, 0])
non_short_prop_idx = np.where(prop_duration <= args.min_length)[0]
keep_idx = [
x for x in keep_idx if x // num_class in non_short_prop_idx
]
# keep_prop_num = len({x//num_class for x in keep_idx})
for k in keep_idx:
cls = k % num_class
prop_idx = k // num_class
if video_id not in dataset_detections[cls]:
dataset_detections[cls][video_id] = np.array([[
rel_prop[prop_idx, 0], rel_prop[prop_idx, 1],
combined_scores[prop_idx, cls], reg_scores[prop_idx, cls, 0],
reg_scores[prop_idx, cls, 1]
]])
else:
dataset_detections[cls][video_id] = np.vstack([
dataset_detections[cls][video_id],
[
rel_prop[prop_idx, 0], rel_prop[prop_idx, 1],
combined_scores[prop_idx, cls],
reg_scores[prop_idx, cls, 0], reg_scores[prop_idx, cls, 1]
]
])
return len(keep_idx)
