def main(argv=sys.argv[1:]):
parser = argparse.ArgumentParser()
parser.add_argument('--input', type=str,
help="Input json file of annotations to remap")
parser.add_argument('--output', type=str,
help="Output json file path for result.")
args = parser.parse_args(argv)
print("Loading source annotation file " + args.input + "...")
annot = rvc_json_helper.load_json(args.input)
anns = []
for a in annot['annotations']:
file_name = a['file_name']
image_id = os.path.splitext(file_name)[0]
for segment in a['segments_info']:
segment['image_id'] = image_id
anns.append(segment)
annot['annotations'] = anns
print("Saving target annotation file "+args.output+"...")
rvc_json_helper.save_json(annot, args.output)
return 0
def main(argv=sys.argv[1:]):
parser = argparse.ArgumentParser()
parser.add_argument('--image_root',
type=str,
default=None,
help="adds image root to each filepath")
parser.add_argument(
'--image_root_rel',
type=str,
default=None,
help=
"adds relative path between output json dir and this directory each filepath"
)
parser.add_argument('--annotations_val',
type=str,
help="Validation annotation file to get categories")
parser.add_argument('--output',
type=str,
help="Output json file path for result.")
args = parser.parse_args(argv)
if not args.image_root_rel is None:
args.image_root_rel, args.image_root = get_relative_path(
args.image_root_rel, args.output)
args.image_root = fix_missing_slash(args.image_root)
args.image_root_rel = fix_missing_slash(args.image_root_rel)
# load annotations
val_annot = rvc_json_helper.load_json(args.annotations_val)
# process images
images = []
image_files = os.listdir(args.image_root_rel)
for file in tqdm(image_files, desc="Processing images ..."):
# get relative filename
file_name = args.image_root.format(file_name=file,
file_name_su=file.split('_')) + file
# get image size
image_path = args.image_root_rel.format(
file_name=file, file_name_su=file.split('_')) + file
width, height = imagesize.get(image_path)
# create image info
img = {
'file_name': file_name,
'id': os.path.splitext(file)[0],
'width': width,
'height': height
}
images.append(img)
# create dataset
test_image_info = {'images': images, 'categories': val_annot['categories']}
# save output
print("Saving target annotation file " + args.output + "...")
rvc_json_helper.save_json(test_image_info, args.output)
return 0
def main(argv=sys.argv[1:]):
parser = argparse.ArgumentParser()
parser.add_argument(
'--join',
type=str,
help=
"Coco json file(s) to join, seperate list entries by semicolon, merging is done in order from first to last"
)
parser.add_argument('--output',
type=str,
help="Output json file path for result.")
args = parser.parse_args(argv)
annot = {}
joint_image_id = 0
for j in args.join.split(';'):
if not os.path.exists(j):
print("Error: file " + j + " not found!")
return 1
print("Loading coco annotation file " + j + "...")
annot_add = rvc_json_helper.load_json(j)
src_info_idx = len(annot_add.get("info", {}).get("datasets", []))
if not 'info' in annot_add:
annot_add["info"] = {"description": " loaded " + j}
if "licenses" in annot_add:
annot_add["info"]["licenses"] = annot_add["licenses"]
if src_info_idx == 0:
annot['info'] = join_info(annot_add)
annot['categories'] = len(annot_add['categories'])
else:
annot['info'] = join_info(annot['info'], annot_add['info'])
if not check_versions_match(annot, annot_add):
print(
"Cannot merge new data to existing file; incompatible formats!"
)
return -3
old_id_to_new = {}
for i in annot_add['images']:
#save original ids for reference
i['ds_id'] = i['id']
i['ds_id'] = src_info_idx # idx from src_info
old_id_to_new[i['id']] = joint_image_id
i['id'] = joint_image_id
joint_image_id += 1
annot['images'] = annot.get('images', []) + annot_add.pop('images')
for a in annot_add['annotations']:
a['image_id'] = old_id_to_new[a['image_id']]
annot['annotations'] = annot.get('annotations',
[]) + annot_add.pop('annotations')
del annot_add
print("Saving result joint annotations to file " + args.output + "...")
rvc_json_helper.save_json(annot, args.output)
return 0
def main(argv=sys.argv[1:]):
parser = argparse.ArgumentParser()
parser.add_argument('--predictions',
type=str,
help="Input json file of predictions to remap")
parser.add_argument(
'--annotations',
type=str,
help="annotations json file used to create the mapping")
parser.add_argument('--mapping',
type=str,
default='./obj_det_mapping.csv',
help="Csv file defining mapping")
parser.add_argument(
'--mapping_row',
type=str,
default=None,
help="Row header from row in csv file to use; default: second row")
parser.add_argument(
'--map_to',
type=str,
default=id,
help="Map to id, freebase_id or any other category type")
parser.add_argument('--void_id',
type=int,
default=0,
help="Void id for labels not found in mapping csv")
parser.add_argument(
'--remove_void',
dest='remove_void',
action='store_true',
help="Only keep boxes with labels that can be remapped")
parser.add_argument('--output',
type=str,
help="Output json file path for result.")
parser.add_argument(
'--reduce_boxable',
dest='reduce_boxable',
action='store_true',
help="Only keep minimum of annotation data needed for boxable training."
)
parser.add_argument(
'--reduce_pano',
dest='reduce_pano',
action='store_true',
help=
"Only keep minimum of annotation data needed for panoptic/instance segm. training."
)
parser.set_defaults(remove_void=False,
reduce_boxable=False,
reduce_pano=False)
args = parser.parse_args(argv)
print("Loading source predictions file " + args.predictions + "...")
pred = rvc_json_helper.load_json(args.predictions)
print("Loading annotation file " + args.annotations + "...")
annot = rvc_json_helper.load_json(args.annotations)
# load pre-defined mapping; first row is target label name, row from --mapping_row
# defines source row for this input data; use semicolon from N->1 mappings
# e.g.
# animal,cat;dog
# will map both cat and dog source labels to the target label animal
with open(args.mapping, newline='') as ifile:
mapping0 = list(csv.reader(ifile))
if args.mapping_row is None:
idx_use = 1
else:
idx_use = mapping0[0].index(args.mapping_row)
trg_labels = {m[0]: m[idx_use] for m in mapping0[1:]}
cats_sorted = list(sorted(trg_labels.keys()))
trg_cats = {
c: {
'supercategory': 'rvc_jls',
'id': id0 + 1,
'name': c
}
for id0, c in enumerate(cats_sorted)
}
src_cats = {c['name'].lower().strip(): c for c in annot['categories']}
trg_to_src = {}
for t, src_all in trg_labels.items():
t_id = trg_cats[t]['id']
for s in src_all.split(';'): #allows multi-to-one mappings
if len(s) == 0:
continue
s = s.lower().strip()
if not s in src_cats: #quick fix for mix of supercategory and name tag in some mvd json file; mix of name/id in oid cat. file
check_cat = "freebase_id" if "freebase_id" in annot[
'categories'][0] else 'supercategory'
pot_supcat = [
c['name'].lower().strip() for c in annot['categories']
if c.get(check_cat, '') == s
]
if len(pot_supcat) == 1:
s = pot_supcat[0]
if not s in src_cats:
print("Warning: Unknown source cat " + s +
" requested. Skipping.")
continue
trg_to_src[t_id] = src_cats[s][args.map_to]
for p in tqdm(pred, desc='Remapping predictions '):
if 'category_id' in p:
p['category_id'] = trg_to_src.get(p['category_id'], args.void_id)
p['bbox'] = [round(box, 2) for box in p['bbox']]
p['score'] = round(p['score'], 4)
if 'segments_info' in p:
for s in p['segments_info']:
s["category_id"] = trg_to_src.get(s['category_id'],
args.void_id)
s['bbox'] = [round(box, 2) for box in s['bbox']]
p['score'] = round(p['score'], 4)
if args.reduce_boxable or args.reduce_pano:
p.pop('segmentation', None)
if args.remove_void:
if 'category_id' in pred[0]:
pred = [p for p in pred if p['category_id'] != args.void_id]
if 'segments_info' in pred[0]:
pred = [[s for s in p if s['category_id'] != args.void_id]
for p in pred]
print("Saving target prediction file " + args.output + "...")
rvc_json_helper.save_json(pred, args.output)
return 0
def main(argv=sys.argv[1:]):
parser = argparse.ArgumentParser()
parser.add_argument('--input', type=str, help="Coco json file to split")
parser.add_argument(
'--split_perc',
type=str,
help=
"Percentage numbers semicolon seperated for splitting. Specify 70;30 for a result with two json files."
)
parser.add_argument(
'--output',
type=str,
default=None,
help=
"Output json file path for result; a _0, _1, _2 etc is added to the basename for each split."
)
args = parser.parse_args(argv)
if args.output is None:
args.output = args.input
print("Loading source coco annotation file " + args.input + "...")
annot = rvc_json_helper.load_json(args.input)
num_imgs = len(annot['images'])
img_to_idx = {}
for im_idx, i in enumerate(annot['images']):
img_to_idx[im_idx] = i
annot['images'] = []
annot_to_idx = {}
for a in annot['annotations']:
annot_to_idx[a['image_id']] = annot_to_idx.get(a['image_id'], []) + [a]
annot['annotations'] = []
all_indices = np.arange(num_imgs)
np.random.shuffle(all_indices)
split_start = 0
for split_idx, split_perc in enumerate(args.split_perc.split(';')):
trg_file = args.output.replace(".json", "_%i.json" % split_idx)
split_num = min(int(float(split_perc) * 0.01 * num_imgs + 0.5),
num_imgs - split_start)
split_idc = list(
sorted(all_indices[split_start:split_start + split_num]))
split_start += split_num
images0 = []
add_image_ids = []
for i in split_idc:
img_entry = img_to_idx.pop(i)
add_image_ids.append(img_entry['id'])
images0.append(img_entry)
annot0 = []
for a in add_image_ids:
annot0 += annot_to_idx.pop(a)
annot_split = annot
annot_split["images"] = images0
annot_split["annotations"] = annot0
print("Saving %i images into coco annotation file " % split_num +
trg_file + "...")
rvc_json_helper.save_json(annot_split, trg_file)
if len(img_to_idx) > 0:
print(
"Warning: A total of %i image entries are in neither of the resulting splitted json files."
% len(img_to_idx))
if len(annot_to_idx) > 0:
print(
"Warning: A total of %i annotation entries are in neither of the resulting splitted json files."
% len(annot_to_idx))
return 0
def main(argv=sys.argv[1:]):
parser = argparse.ArgumentParser()
parser.add_argument('--input', type=str,
help="Input json file of annotations to remap")
parser.add_argument('--mapping', type=str, default='./obj_det_mapping.csv',
help="Csv file defining mapping")
parser.add_argument('--mapping_row', type=str, default=None,
help="Row header from row in csv file to use; default: second row")
parser.add_argument('--image_root', type=str, default=None,
help="adds image root to each filepath")
parser.add_argument('--image_root_rel', type=str, default=None,
help="adds relative path between output json dir and this directory each filepath")
parser.add_argument('--annotation_root', type=str, default=None,
help="adds annotation mask root to each filepath")
parser.add_argument('--annotation_root_rel', type=str, default=None,
help="adds relative path between output json dir and annotation mask directory each filepath")
parser.add_argument('--void_id', type=int, default=0,
help="Void id for labels not found in mapping csv")
parser.add_argument('--output', type=str,
help="Output json file path for result.")
parser.add_argument('--reduce_boxable', dest='reduce_boxable', action='store_true',
help="Only keep minimum of annotation data needed for boxable training.")
parser.add_argument('--reduce_pano', dest='reduce_pano', action='store_true',
help="Only keep minimum of annotation data needed for panoptic/instance segm. training.")
parser.set_defaults(do_merging=False, reduce_boxable=False, reduce_pano=False)
args = parser.parse_args(argv)
if not args.image_root_rel is None:
args.image_root_rel, args.image_root = get_relative_path(args.image_root_rel, args.output)
if not args.annotation_root_rel is None:
args.annotation_root_rel, args.annotation_root = get_relative_path(args.annotation_root_rel, args.output)
fix_missing_slash(args.image_root)
fix_missing_slash(args.annotation_root)
print("Loading source annotation file " + args.input + "...")
annot = rvc_json_helper.load_json(args.input)
# load pre-defined mapping; first row is target label name, row from --mapping_row
# defines source row for this input data; use semicolon from N->1 mappings
# e.g.
# animal,cat;dog
# will map both cat and dog source labels to the target label animal
with open(args.mapping, newline='') as ifile:
mapping0 = list(csv.reader(ifile))
if args.mapping_row is None:
idx_use = 1
else:
idx_use = mapping0[0].index(args.mapping_row)
trg_labels = {m[0]:m[idx_use] for m in mapping0[1:]}
cats_sorted = list(sorted(trg_labels.keys()))
trg_cats = {c:{'supercategory':'rvc_jls', 'id': id0+1, 'name':c} for id0, c in enumerate(cats_sorted)}
src_cats = {c['name'].lower().strip():c for c in annot['categories']}
src_to_trg = {}
for t, src_all in trg_labels.items():
t_id = trg_cats[t]['id']
for s in src_all.split(';'): #allows multi-to-one mappings
if len(s) == 0:
continue
s = s.lower().strip()
if not s in src_cats: #quick fix for mix of supercategory and name tag in some mvs json file; mix of name/id in oid cat. file
check_cat = "freebase_id" if "freebase_id" in annot['categories'][0] else 'supercategory'
pot_supcat = [c['name'].lower().strip() for c in annot['categories'] if c.get(check_cat,'') == s]
if len(pot_supcat) == 1:
s = pot_supcat[0]
if not s in src_cats:
print("Warning: Unknown source cat "+s+" requested. Skipping.")
continue
src_to_trg[src_cats[s]['id']]= t_id
if 'categories' in annot:
annot['categories'] = [trg_cats[c] for c in cats_sorted]
if not args.image_root is None:
if args.image_root[-1] != '/' and args.image_root[-1] != '\\':
args.image_root += '/'
for i in annot['images']:
i['file_name'] = args.image_root.format(file_name=i['file_name'], file_name_su = i['file_name'].split('_')) + i['file_name']
if args.reduce_pano or args.reduce_boxable:
reduce_size_entries = ["date_captured","coco_url","url","flickr_url"]
for i in annot['images']:
for e in reduce_size_entries:
i.pop(e,None)
if 'annotations' in annot:
for a in tqdm.tqdm(annot['annotations'], desc='Remapping annotations '):
if 'category_id' in a:
a['category_id'] = src_to_trg.get(a['category_id'],args.void_id)
if 'segments_info' in a:
for s in a['segments_info']:
s["category_id"] = src_to_trg.get(s['category_id'],args.void_id)
if args.reduce_boxable or args.reduce_pano:
a.pop('segmentation',None)
if not args.annotation_root is None and 'file_name' in a:
a['file_name'] = args.annotation_root.format(file_name=a['file_name'], file_name_su = a['file_name'].split('_')) + a['file_name']
print("Saving target annotation file "+args.output+"...")
rvc_json_helper.save_json(annot, args.output)
return 0