def generate_iou_data(queries,
path,
if_data,
use_geometric=False,
iqd_params=None,
pred_weight=None):
for query_id, query in tqdm(enumerate(queries), total=len(queries)):
object_id_list = []
iou_list = []
tp_simple = data_utils.get_partial_query_matches(
if_data.vg_data, queries)
for image_id in tqdm(tp_simple[query_id]):
if iqd_params is None:
iqd_params = {}
iqd = ImageQueryData(query,
query_id,
image_id,
if_data,
use_geometric=use_geometric,
pred_weight=pred_weight,
**iqd_params)
iqd.compute_potential_data()
object_id_list.extend(
[iqd.model_sub_bbox_id, iqd.model_obj_bbox_id])
iou_list.extend([iqd.sub_model_iou, iqd.obj_model_iou])
iou_name = 'q{:03d}_iou_values.csv'.format(query_id)
iou_path = os.path.join(path, iou_name)
with open(iou_path, 'wb') as f:
csv_writer = csv.writer(f)
csv_writer.writerow(('image_ix', 'obj_ix', 'iou'))
for image_id, (object_id,
iou) in enumerate(zip(object_id_list, iou_list)):
csv_writer.writerow((image_id, object_id, iou))
def save_query_texts(queries, if_data, path):
"""Save the texts of the given queries to a file."""
tp_simple = data_utils.get_partial_query_matches(if_data.vg_data, queries)
with open(path, 'w') as f:
for index, (query, image_index) in enumerate(zip(queries, tp_simple)):
iqd = query_viz.ImageQueryData(query, index, 0, if_data)
f.write('[{}] (count: {}): {}\n'.format(index, len(image_index),
iqd.get_query_text()))
def iou_check(queries, if_data):
tp_simple = data_utils.get_partial_query_matches(if_data.vg_data, queries)
energy_path = os.path.join(out_path, 'query_ious/')
geom_energy_path = os.path.join(out_path, 'query_ious_geom/')
data_simple = [(energy_path, 'factor graph'),
(geom_energy_path, 'geometric mean')]
data_utils.get_iou_recall_values(data_simple,
tp_simple,
len(if_data.vg_data),
show_plot=True)
def export_images(queries, if_data, false_negs, path, json_data):
"""Save the counts of the given queries to a file."""
tp_simple = data_utils.get_partial_query_matches(if_data.vg_data, queries)
for query_index, image_index in false_negs:
tp_simple[query_index].append(image_index)
pairs = enumerate(zip(queries, tp_simple))
for index, (query, image_indexes) in tqdm(pairs,
desc='queries',
total=len(queries)):
iqd = ImageQueryData(query, index, 0, if_data)
query_dir = iqd.get_query_text().replace(' ', '_')
query_path = os.path.join(path, query_dir)
if not os.path.exists(query_path):
os.makedirs(query_path)
for image_index in image_indexes:
if_data.configure(image_index, None)
shutil.copy(if_data.image_filename, query_path)
image_name = os.path.basename(if_data.image_filename)
json_name = os.path.splitext(image_name)[0] + '.json'
with open(os.path.join(query_path, json_name), 'w') as f:
json.dump(json_data[image_index], f)
neg_path = os.path.join(path, 'negative_images')
if not os.path.exists(neg_path):
os.makedirs(neg_path)
full_negs = [
index for index in range(len(if_data.vg_data))
if all([index not in image_index for image_index in tp_simple])
]
for image_index in full_negs:
if_data.configure(image_index, None)
shutil.copy(if_data.image_filename, neg_path)
image_name = os.path.basename(if_data.image_filename)
json_name = os.path.splitext(image_name)[0] + '.json'
with open(os.path.join(neg_path, json_name), 'w') as f:
json.dump(json_data[image_index], f)
def recall_check(queries, if_data, false_negs, situate_recalls):
"""Compute recalls for a given set of queries and plot."""
tp_simple = data_utils.get_partial_query_matches(if_data.vg_data, queries)
for query_index, image_index in false_negs:
tp_simple[query_index].append(image_index)
energy_path = os.path.join(out_path, 'query_energies_psu/')
geom_energy_path = os.path.join(out_path, 'query_energies_geom_psu/')
true_geom_energy_path = os.path.join(out_path,
'query_energies_true_geom_psu/')
weighted_energy_path = os.path.join(out_path,
'query_energies_weighted_psu/')
rcnn_energy_path = os.path.join(out_path, 'query_energies_rcnn_psu/')
data_simple = [(energy_path, 'vanilla IRSG'),
(geom_energy_path, 'geometric mean on potentials'),
(true_geom_energy_path, 'true geometric mean'),
(weighted_energy_path, 'weighted IRSG'),
(rcnn_energy_path, 'RCNN-weighted IRSG')]
data_utils.get_single_image_recall_values(data_simple,
tp_simple,
len(if_data.vg_data),
show_plot=True,
situate_recalls=situate_recalls)
def save_query_counts(queries, if_data, false_negs, path):
"""Save the counts of the given queries to a file."""
tp_simple = data_utils.get_partial_query_matches(if_data.vg_data, queries)
for query_index, image_index in false_negs:
tp_simple[query_index].append(image_index)
with open(path, 'w') as f:
for index, (query, image_index) in enumerate(zip(queries, tp_simple)):
iqd = ImageQueryData(query, index, 0, if_data)
f.write('{} ({} pos, {} neg)\n'.format(
iqd.get_query_text(), len(image_index),
len(if_data.vg_data) - len(image_index)))
f.write('positives: {}\n'.format(image_index))
duplicates = list(
set([
index for index in image_index
if image_index.count(index) > 1
]))
f.write('duplicates: {}\n\n'.format(duplicates))
full_negs = [
index for index in range(len(if_data.vg_data))
if all([index not in image_index for image_index in tp_simple])
]
f.write('negatives for ALL queries: {}'.format(len(full_negs)))
parts = [part.replace('_', ' ') for part in line.split()]
text_parts, gen_func = parts[:-1], gen_dict[parts[-1]]
query_struct.annotations = gen_func(
*text_parts, use_attrs=use_attrs)
queries.append(query_struct)
return queries
if __name__ == '__main__':
_, _, _, _, _, if_data = dp.get_all_data(
'stanford', split='test', use_csv=True)
path = os.path.join(data_path, 'queries.txt')
queries = generate_queries_from_file(path)
false_neg_path = os.path.join(data_path, 'false_negs.csv')
false_negs = data_utils.get_false_negs(false_neg_path)
tp_data_pos = data_utils.get_partial_query_matches(
if_data.vg_data, queries)
for query_index, image_index in false_negs:
tp_data_pos[query_index].append(image_index)
tp_data_neg = generate_tp_neg(tp_data_pos, len(queries),
len(if_data.vg_data), NEGS_PER_QUERY)
generate_all_query_plots(queries, tp_data_pos, tp_data_neg,
if_data, condition_gmm=True,
visualize_gmm=False, false_negs=false_negs)
generate_all_query_plots(queries, tp_data_pos, tp_data_neg,
if_data, condition_gmm=True,
visualize_gmm=False, use_geometric=True,
suffix='true_geom', false_negs=false_negs)
generate_all_query_plots(queries, tp_data_pos, tp_data_neg,
if_data, condition_gmm=True,
visualize_gmm=False, use_alt_geom=True,
suffix='geom', false_negs=false_negs)