def main():
args = parse_opts()
set_random_seed(args.random_seed)
logger_name = "retrieval_data"
retrieval_dataset_name_suffix = "-retrieval"
logger = setup_logger(logger_name, None)
data_path = get_data_path()
script_path = os.path.expanduser(os.path.dirname(os.path.abspath(__file__)))
target_path = os.path.join(script_path, "cnnimageretrieval-pytorch", "data")
mkdir(target_path)
dataset_train = build_dataset_by_name(data_path, args.dataset_train,
eval_scale=args.dataset_train_scale,
logger_prefix=logger_name)
retrieval_dataset_train_name = dataset_train.get_name() + retrieval_dataset_name_suffix
dataset_val = build_dataset_by_name(data_path, args.dataset_val,
eval_scale=args.dataset_val_scale,
logger_prefix=logger_name)
retrieval_dataset_val_name = dataset_val.get_name() + retrieval_dataset_name_suffix
datasets_test = []
retrieval_dataset_test_names = []
if args.datasets_test:
if len(args.datasets_test_scale) == 1:
datasets_test_scale = args.datasets_test_scale * len(args.datasets_test)
else:
datasets_test_scale = args.datasets_test_scale
assert len(args.datasets_test) == len(datasets_test_scale), "Arg datasets-test-scale should be of len 1 or of len equal to the len of datasets-test"
for dataset_name, scale in zip(args.datasets_test, datasets_test_scale):
dataset = build_dataset_by_name(data_path, dataset_name,
eval_scale=scale,
logger_prefix=logger_name)
retrieval_dataset_test_names.append(dataset.get_name() + retrieval_dataset_name_suffix)
datasets_test.append(dataset)
# create dataset
if args.num_random_crops_per_image > 0:
crop_suffix = f"-rndCropPerImage{args.num_random_crops_per_image}"
retrieval_dataset_train_name = retrieval_dataset_train_name + crop_suffix
retrieval_dataset_val_name = retrieval_dataset_val_name + crop_suffix
retrieval_dataset_test_names = [name + crop_suffix for name in retrieval_dataset_test_names]
prepare_dataset(target_path,
retrieval_dataset_train_name, dataset_train,
retrieval_dataset_val_name, dataset_val,
args.iou_pos_threshold, args.iou_neg_threshold,
args.num_queries_image_to_image,
logger,
retrieval_dataset_test_names=retrieval_dataset_test_names, datasets_test=datasets_test,
num_random_crops_per_image=args.num_random_crops_per_image)
def build_os2d_dataset_by_name(name, data_path=None):
if data_path is None:
data_path = cfg.DATA_DIR
dataset = build_dataset_by_name(data_path,
name,
eval_scale=None,
cache_images=False)
return Os2dDataset(dataset)
def main():
target_object_size = 240
data_path = get_data_path()
logger = setup_logger("get_dataset_scales", None)
for name in DATASET_LIST:
dataset = build_dataset_by_name(data_path, name, eval_scale=None, logger_prefix="get_dataset_scales")
image_sizes_by_id = get_image_sizes(dataset)
average_size, median, q10, q90 = compute_average_object_size(dataset.gtboxframe, image_sizes_by_id)
print("Average size of object = {0:0.2f} for image size = {1}".format(average_size, dataset.image_size))
print("Median = {0:0.2f}, q10 = {1:0.2f}, q90 = {2:0.2f}".format(median, q10, q90))
print("To get objects to size {0}, images should be of size {1:d}".format(target_object_size, int(dataset.image_size * target_object_size / median)))
def build_detection_dataset_by_name(data_path,
name,
transforms,
class_ids=None,
cache_images=False,
ignore_labels=False):
dataset_os2d = build_dataset_by_name(data_path,
name,
cache_images=cache_images,
no_image_reading=False)
dataset = DatasetDetection(dataset_os2d,
transforms,
class_ids=class_ids,
ignore_labels=ignore_labels)
return dataset
def main():
args = parse_opts()
set_random_seed(args.random_seed)
crop_suffix = f"-rndCropPerImage{args.num_random_crops_per_image}"
logger_name = "retrieval_data"
retrieval_dataset_name_suffix = "-retrieval"
logger = setup_logger(logger_name, None)
data_path = get_data_path()
script_path = os.path.expanduser(os.path.dirname(
os.path.abspath(__file__)))
target_path = os.path.join(script_path, "cnnimageretrieval-pytorch",
"data")
mkdir(target_path)
dataset_train = build_dataset_by_name(data_path,
args.dataset_train,
eval_scale=args.dataset_train_scale,
logger_prefix=logger_name)
retrieval_dataset_train_name = dataset_train.get_name(
) + retrieval_dataset_name_suffix
dataset_val = build_dataset_by_name(data_path,
args.dataset_val,
eval_scale=args.dataset_val_scale,
logger_prefix=logger_name)
retrieval_dataset_val_name = dataset_val.get_name(
) + retrieval_dataset_name_suffix
if args.datasets_test:
if len(args.datasets_test_scale) == 1:
datasets_test_scale = args.datasets_test_scale * len(
args.datasets_test)
else:
datasets_test_scale = args.datasets_test_scale
assert len(args.datasets_test) == len(
datasets_test_scale
), "Arg datasets-test-scale should ne of len 1 or of len equal to the len of datasets-test"
datasets_test = []
retrieval_dataset_test_names = []
for dataset_name, scale in zip(args.datasets_test,
datasets_test_scale):
dataset = build_dataset_by_name(data_path,
dataset_name,
eval_scale=scale,
logger_prefix=logger_name)
retrieval_dataset_test_names.append(dataset.get_name() +
retrieval_dataset_name_suffix)
datasets_test.append(dataset)
# prepare data images for train and val
tgt_image_path_trainval = os.path.join(target_path, "train",
retrieval_dataset_train_name, "ims")
mkdir(tgt_image_path_trainval)
logger.info(
f"Train set {retrieval_dataset_train_name} with no random crops")
db_images_train = save_cropped_boxes(dataset_train,
tgt_image_path_trainval,
extension="")
# create val subset: add all boxes from images that have at least one validation box (can add some boxes from train as distractors)
logger.info(f"Val set {retrieval_dataset_val_name} with no random crops")
db_images_val = save_cropped_boxes(dataset_val,
tgt_image_path_trainval,
extension="")
# prepare data images for trainval with crops
tgt_image_path_trainval_randcrops = os.path.join(
target_path, "train", retrieval_dataset_train_name + crop_suffix,
"ims")
mkdir(tgt_image_path_trainval_randcrops)
logger.info(
f"Train set {retrieval_dataset_train_name} with {args.num_random_crops_per_image} crops per image"
)
db_images_train_randomCrops = save_cropped_boxes(
dataset_train,
tgt_image_path_trainval_randcrops,
extension="",
num_random_crops_per_image=args.num_random_crops_per_image)
# create val subset: add all boxes from images that have at least one validation box (can add some boxes from train as distractors)
logger.info(
f"Val set {retrieval_dataset_val_name} with {args.num_random_crops_per_image} crops per image"
)
db_images_val_randomCrops = save_cropped_boxes(
dataset_val,
tgt_image_path_trainval_randcrops,
extension="",
num_random_crops_per_image=args.num_random_crops_per_image)
# prepare data images for test
dbs_images_test = {}
if datasets_test:
for dataset_test, dataset_name in zip(datasets_test,
retrieval_dataset_test_names):
tgt_image_path_test = os.path.join(
target_path, "test", dataset_name, "jpg"
) # the folder name should be always "test" - from cirtorch
mkdir(tgt_image_path_test)
logger.info(f"Eval dataset: {dataset_name}")
dbs_images_test[dataset_name] = save_cropped_boxes(
dataset_test, tgt_image_path_test)
# prepare data images for test with random crops
tgt_image_path_test = os.path.join(
target_path, "test", dataset_name + crop_suffix, "jpg"
) # the folder name should be always "test" - from cirtorch
mkdir(tgt_image_path_test)
logger.info(f"Eval dataset: {dataset_name + crop_suffix}")
dbs_images_test[dataset_name + crop_suffix] = save_cropped_boxes(
dataset_test,
tgt_image_path_test,
num_random_crops_per_image=args.num_random_crops_per_image)
# save GT images from train
db_classes_train = save_class_images(
dataset_train,
os.path.join(target_path, "train", retrieval_dataset_train_name,
"ims"),
extension="")
db_classes_train_randomCrops = save_class_images(
dataset_train,
os.path.join(target_path, "train",
retrieval_dataset_train_name + crop_suffix, "ims"),
extension="")
# save GT images from val
db_classes_val = save_class_images(
dataset_val,
os.path.join(target_path, "train", retrieval_dataset_train_name,
"ims"),
extension="")
db_classes_val_randomCrops = save_class_images(
dataset_val,
os.path.join(target_path, "train",
retrieval_dataset_train_name + crop_suffix, "ims"),
extension="")
# save GT images for testing
dbs_classes_test = {}
if args.datasets_test:
for dataset_test, dataset_name in zip(datasets_test,
retrieval_dataset_test_names):
dbs_classes_test[dataset_name] = save_class_images(
dataset_test,
os.path.join(target_path, "test", dataset_name, "jpg"))
dbs_classes_test[dataset_name + crop_suffix] = save_class_images(
dataset_test,
os.path.join(target_path, "test", dataset_name + crop_suffix,
"jpg"))
# merge databases
logger.info(
f"Processing trainval set from {retrieval_dataset_train_name} and {retrieval_dataset_val_name}"
)
db_train = create_train_database_queries(
db_images_train,
db_classes_train,
iou_pos_threshold=args.iou_pos_threshold,
iou_neg_threshold=args.iou_neg_threshold,
logger=logger)
db_val = create_train_database_queries(
db_images_val,
db_classes_val,
iou_pos_threshold=args.iou_pos_threshold,
iou_neg_threshold=args.iou_neg_threshold,
logger=logger)
logger.info(
f"Processing trainval set from {retrieval_dataset_train_name} and {retrieval_dataset_val_name} with {args.num_random_crops_per_image} random crops"
)
db_train_randomCrops = create_train_database_queries(
db_images_train_randomCrops,
db_classes_train_randomCrops,
iou_pos_threshold=args.iou_pos_threshold,
iou_neg_threshold=args.iou_neg_threshold,
logger=logger)
db_val_randomCrops = create_train_database_queries(
db_images_val_randomCrops,
db_classes_val_randomCrops,
iou_pos_threshold=args.iou_pos_threshold,
iou_neg_threshold=args.iou_neg_threshold,
logger=logger)
dbs_test = {}
if args.datasets_test:
for dataset_name in retrieval_dataset_test_names:
logger.info(
f"Processing test set {dataset_name} with {args.num_random_crops_per_image} random crops"
)
dbs_test[dataset_name] = create_test_database_queries(
dbs_images_test[dataset_name],
dbs_classes_test[dataset_name],
iou_pos_threshold=args.iou_pos_threshold,
iou_neg_threshold=args.iou_neg_threshold,
logger=logger)
logger.info(f"Processing test set {dataset_name + crop_suffix}")
dbs_test[dataset_name +
crop_suffix] = create_test_database_queries(
dbs_images_test[dataset_name + crop_suffix],
dbs_classes_test[dataset_name + crop_suffix],
iou_pos_threshold=args.iou_pos_threshold,
iou_neg_threshold=args.iou_neg_threshold,
logger=logger)
# save trainval to disk
db_trainval = {"train": db_train, "val": db_val}
db_fn = os.path.join(
os.path.join(target_path, "train", retrieval_dataset_train_name),
f"{retrieval_dataset_train_name}.pkl")
with open(db_fn, "wb") as f:
pickle.dump(db_trainval, f)
# save train separately for whitening
db_fn = os.path.join(
os.path.join(target_path, "train", retrieval_dataset_train_name),
f"{retrieval_dataset_train_name}-whiten.pkl")
with open(db_fn, "wb") as f:
pickle.dump(db_train, f)
# save trainval with random crops to disk
db_trainval_randomCrops = {
"train": db_train_randomCrops,
"val": db_val_randomCrops
}
db_fn = os.path.join(
os.path.join(target_path, "train",
retrieval_dataset_train_name + crop_suffix),
f"{retrieval_dataset_train_name}{crop_suffix}.pkl")
with open(db_fn, "wb") as f:
pickle.dump(db_trainval_randomCrops, f)
db_fn = os.path.join(
os.path.join(target_path, "train",
retrieval_dataset_train_name + crop_suffix),
f"{retrieval_dataset_train_name}{crop_suffix}-whiten.pkl")
with open(db_fn, "wb") as f:
pickle.dump(db_train_randomCrops, f)
# save test to disk
if args.datasets_test:
for dataset_name in retrieval_dataset_test_names:
db_fn = os.path.join(
os.path.join(target_path, "test", dataset_name),
f"gnd_{dataset_name}.pkl")
with open(db_fn, "wb") as f:
pickle.dump(dbs_test[dataset_name], f)
# save test with random crops to disk
db_fn = os.path.join(
os.path.join(target_path, "test", dataset_name + crop_suffix),
f"gnd_{dataset_name}{crop_suffix}.pkl")
with open(db_fn, "wb") as f:
pickle.dump(dbs_test[dataset_name + crop_suffix], f)
