see-through/training/scripts/data_pipeline.py
Miaomiao Li 9ded69536e public: release training scripts, configs, and data pipeline (V3)
Co-Authored-By: dmMaze <beneathlimbo@gmail.com>
2026-04-14 15:16:51 +08:00

1984 lines
No EOL
80 KiB
Python

import os
import random
import os.path as osp
import numpy as np
from pathlib import Path
import shutil
import sys
sys.path.append(osp.dirname(osp.dirname(osp.abspath(__file__))))
from PIL import Image
from tqdm import tqdm
import click
import cv2
from utils.cv import mask2rle, rle2mask, mask_xyxy
from utils.io_utils import load_exec_list, find_all_files_recursive, find_all_files_with_name, pil_ensure_rgb, imglist2imgrid, imread, imwrite, json2dict, save_tmp_img, dict2json
from utils.sampler import NameSampler
from utils.visualize import visualize_segs, visualize_segs_with_labels, visualize_pos_keypoints
from live2d.scrap_model import Live2DScrapModel, VALID_BODY_PARTS_V1, VALID_BODY_PARTS_V2, compose_mask_from_drawables, init_drawable_visible_map, load_detected_character, load_pos_estimation
@click.group()
def cli():
"""live2d data processing related scripts.
"""
def get_unique_render_lst(exec_list):
unique_lst = []
processed_models = set()
unique_src_to_models = dict()
for p in tqdm(exec_list):
modeld = osp.dirname(p)
if modeld not in processed_models:
processed_models.add(modeld)
else:
continue
plist = sub_render_parts([p])
mlist = [Live2DScrapModel(p, skip_load=True) for p in plist]
for m in mlist:
m.init_drawables()
unique_mlist = [mlist[4]]
for m in mlist:
is_unique = True
mklist = list(m.did2drawable.keys())
mklist.sort()
for um in unique_mlist:
umklist = list(um.did2drawable.keys())
umklist.sort()
if mklist == umklist:
srcp = um.directory
is_unique = False
break
tgtp = m.directory
if is_unique:
unique_mlist.append(m)
srcp = m.directory
if srcp not in unique_src_to_models:
unique_src_to_models[srcp] = []
unique_src_to_models[srcp].append(tgtp)
unique_mlist = [m.directory for m in unique_mlist]
unique_lst += unique_mlist
return unique_lst, unique_src_to_models
@cli.command('get_tgt_list')
@click.option('--src_dir')
@click.option('--savep', default=None)
def get_tgt_list(src_dir, savep):
if savep is None:
savep = osp.join('workspace/datasets', osp.basename(src_dir) + '.txt')
valid_list = []
for f in find_all_files_recursive(src_dir, ext={'.json'}):
tgtf = f.rstrip('.json') + '.png'
if osp.exists(tgtf):
valid_list.append(tgtf)
print(len(valid_list))
with open(savep, 'w', encoding='utf8') as f:
f.write('\n'.join(valid_list))
@cli.command('get_png_list')
@click.option('--src_dir')
@click.option('--savep', default=None)
def get_png_list(src_dir, savep):
if savep is None:
savep = osp.join('workspace/datasets', osp.basename(src_dir) + '.txt')
valid_list = []
for f in find_all_files_recursive(src_dir, ext={'.png'}):
valid_list.append(f)
print(len(valid_list))
with open(savep, 'w', encoding='utf8') as f:
f.write('\n'.join(valid_list))
@cli.command('check_unique_rst')
@click.option('--exec_list')
@click.option('--savep', default=None)
def check_unique_rst(exec_list, savep):
if savep is None:
savep = exec_list
exec_list = load_exec_list(exec_list)
exec_list, unique_src_to_models = get_unique_render_lst(exec_list)
print(len(exec_list))
with open(savep, 'w', encoding='utf8') as f:
f.write('\n'.join(exec_list))
dict2json(unique_src_to_models, savep + '.json')
@cli.command('compress_live2d')
@click.option('--src_dir')
@click.option('--save_dir')
@click.option('--ext', default='.jxl')
@click.option('--disable_crop', is_flag=True, default=False)
def compress_live2d(src_dir, save_dir, ext, disable_crop):
src_dir = osp.normpath(src_dir)
model_final_list = find_all_files_with_name(src_dir, 'final')
crop = not disable_crop
if save_dir is None:
save_dir = src_dir + f'_{ext}'
if crop:
save_dir += '_crop'
save_dir = osp.normpath(save_dir)
os.makedirs(save_dir, exist_ok=True)
ndir_leading = len(src_dir.split(os.path.sep))
for model_f in tqdm(model_final_list, desc=f'saving to {save_dir}'):
model_dir = osp.dirname(model_f)
model_save_dir = model_dir.split(os.path.sep)[ndir_leading:]
model = Live2DScrapModel(model_dir, crop_to_final=crop, pad_to_square=False)
model.save_model_to(osp.join(save_dir, *model_save_dir),
crop_to_final=crop, img_ext=ext)
@cli.command('build_live2d_exec_list')
@click.option('--srcd')
@click.option('--save_dir', default=None)
@click.option('--filter_p', default=None)
@click.option('--target_fno', default=-1)
@click.option('--num_chunk', default=-1)
@click.option('--save_name', default='exec_list')
def build_live2d_exec_list(srcd, save_dir, filter_p, target_fno, num_chunk, save_name):
exec_list = find_all_files_with_name(srcd, name='final', exclude_suffix=True)
tgt_list = []
filter_set = set()
if filter_p is not None:
filter_set = set(load_exec_list(filter_p))
for d in exec_list:
if d in filter_set or osp.dirname(d) in filter_set:
continue
dname = osp.basename(osp.dirname(d))
if target_fno > 0:
fno = dname.split('-')[-1]
if not fno.isdigit():
print(f'{d} is not a valid path')
continue
fno = int(fno)
if fno == target_fno:
tgt_list.append(d)
else:
tgt_list.append(d)
random.shuffle(tgt_list)
print(f'num samples: {len(tgt_list)}')
if save_dir is None:
save_dir = srcd
with open(osp.join(save_dir, f'{save_name}.txt'), 'w', encoding='utf8') as f:
f.write('\n'.join(tgt_list))
if num_chunk > 0:
world_size = num_chunk
for ii in range(world_size):
t = load_exec_list(tgt_list, ii, world_size=world_size)
with open(osp.join(save_dir, f'{save_name}{ii}.txt'), 'w', encoding='utf8') as f:
f.write('\n'.join(t))
print(f'chunk {ii} num samples: {len(t)}')
@cli.command('build_exec_list')
@click.option('--srcd')
@click.option('--exts', default=None)
@click.option('--save_dir', default=None)
@click.option('--num_chunk', default=-1)
@click.option('--save_name', default='exec_list')
def build_exec_list(srcd, exts, save_dir, num_chunk, save_name):
exec_list = []
exts = exts.split(',')
for s in srcd.split(','):
exec_list += find_all_files_recursive(s, ext=exts)
print(f'found {len(exec_list)} samples for {exts}')
tgt_list = exec_list
random.shuffle(tgt_list)
if save_dir is None:
save_dir = srcd
with open(osp.join(save_dir, f'{save_name}.txt'), 'w', encoding='utf8') as f:
f.write('\n'.join(tgt_list))
if num_chunk > 0:
world_size = num_chunk
for ii in range(world_size):
t = load_exec_list(tgt_list, ii, world_size=world_size)
with open(osp.join(save_dir, f'{save_name}{ii}.txt'), 'w', encoding='utf8') as f:
f.write('\n'.join(t))
print(f'chunk {ii} num samples: {len(t)}')
def sub_render_parts(exec_list):
new_exec_list = []
for d in exec_list:
d = '-'.join(d.split('-')[:-1])
for ii in range(9):
md = d + f'-{ii}'
if osp.exists(md):
new_exec_list.append(md)
return new_exec_list
def assign_masks_to_points(points, mask_list, distance_thr=1.0):
def _coord_stats(cx):
x1, x2 = np.min(cx), np.max(cx)
cent_x = np.round(np.mean(cx))
return x1, x2, cent_x
mask_asignment = {}
n_points = len(points)
n_mask = len(mask_list)
if n_points < 1 or n_mask < 1:
return mask_asignment
for ii in range(n_points):
mask_asignment[ii] = {'mask': None, 'distance': [], 'mask_idx': None}
h, w = mask_list[0].shape[:2]
dist_max = h * w
for m in mask_list:
coords = np.where(m)
if len(coords[0]) == 0:
for pi in range(n_points):
mask_asignment[pi]['distance'].append(dist_max)
continue
x1, x2, cx = _coord_stats(coords[1])
y1, y2, cy = _coord_stats(coords[0])
diag = np.sqrt((x1 - x2) ** 2 + (y1 - y2) ** 2)
for pi, pnt in enumerate(points):
dist = np.linalg.norm(pnt - np.array([cx, cy]))
if dist > distance_thr * diag:
dist = dist_max
mask_asignment[pi]['distance'].append(dist)
for pi in range(n_points):
mask_asignment[pi]['distance'] = np.array(mask_asignment[pi]['distance'])
not_assigned = list(range(n_points))
while len(not_assigned) > 0:
pi_dist_pair = [(pi, np.min(mask_asignment[pi]['distance']), np.argmin(mask_asignment[pi]['distance'])) for pi in not_assigned]
pi_dist_pair.sort(key=lambda x: x[1])
m_pi, m_dist, m_mid = pi_dist_pair.pop(0)
if m_dist < dist_max:
mask_asignment[m_pi]['mask'] = mask_list[m_mid]
mask_asignment[m_pi]['mask_idx'] = m_mid
else:
break
not_assigned = [p[0] for p in pi_dist_pair if p[1] < dist_max]
for pi in not_assigned:
mask_asignment[pi]['distance'][m_mid] = dist_max
pass
return mask_asignment
def try_assign_sam_mask(lmodel: Live2DScrapModel, mask_list, body_part_tag, check_points=None,
mask_ioa_thr=0.4, valid_face_ids=None, skip_existed_bodypart=True,
exclude_mask=None, bbox_constraint=None, y2_max=None):
if exclude_mask is not None:
mask_list = [m for m in mask_list if not np.any(m & exclude_mask)]
if bbox_constraint is not None or y2_max is not None:
_mask_list = []
for m in mask_list:
coords = np.where(m)
mx1, mx2 = np.min(coords[1]), np.max(coords[1])
my1, my2 = np.min(coords[0]), np.max(coords[0])
if bbox_constraint is not None:
if my1 < bbox_constraint[1] or my2 > bbox_constraint[3] or mx1 < bbox_constraint[0] or mx2 > bbox_constraint[2]:
continue
if y2_max is not None and my2 > y2_max:
continue
_mask_list.append(m)
mask_list = _mask_list
if check_points is not None:
msk_assignment = assign_masks_to_points(check_points, mask_list)
msk = [m['mask'] for m in msk_assignment.values() if m['mask'] is not None]
else:
msk = mask_list
if len(msk) < 1:
return None, None
if len(msk) > 1:
msk = np.logical_or.reduce(np.stack(msk), axis=0)
else:
msk = msk[0]
msk_assigned = False
if valid_face_ids is None:
valid_face_ids = set()
msk_area = msk.sum() + 1e-6
ioa_lst = {}
for ii, d in enumerate(lmodel.drawables):
if d.final_visible_area < 1:
continue
if d.face_part_id is not None and d.face_part_id not in valid_face_ids:
continue
if skip_existed_bodypart and d.body_part_tag is not None:
continue
x1, y1, x2, y2 = d.xyxy
mask_sum = np.sum(msk[y1: y2, x1: x2] & d.final_visible_mask)
ioa = mask_sum / d.final_visible_area
ios = mask_sum / msk_area
ioa_lst[ii] = {'ioa': ioa, 'ios': ios}
if ioa > mask_ioa_thr or ios > 1.:
d.body_part_tag = body_part_tag
msk_assigned = True
# if not msk_assigned:
# max_ioa = -1
# assigned_drawable = None
# for ii, d in enumerate(lmodel.drawables):
# if d.final_visible_area < 1:
# continue
# if d.face_part_id is not None and d.face_part_id not in valid_face_ids:
# continue
# if skip_existed_bodypart and d.body_part_tag is not None:
# continue
# ioa, ios = ioa_lst[ii]['ioa'], ioa_lst[ii]['ios']
# if ios > 0.5:
# assigned_drawable = d
# max_ioa = ioa
# if assigned_drawable is not None:
return msk, msk_assigned
def mask_cover_pos(mask, pos_list, pos_ids, mode='any', xshift=0, yshift=0):
if isinstance(pos_ids, (int, np.ScalarType)):
pos_ids = [pos_ids]
h, w = mask.shape[:2]
# if mode == 'any':
for pi in pos_ids:
pi = pos_list[pi]
py, px = pi[1] + yshift, pi[0] + xshift
covered = py < h and py >= 0 and px < w and px >= 0
if covered and mask[py, px] > 0:
if mode == 'any':
return True
else:
return False
if mode == 'any':
return True
else:
return False
def mask_line_sample(mask, line_start, line_end, divide_long_side=False):
x1, y1 = line_start
x2, y2 = line_end
lh, lw = abs(y2 - y1), abs(x2 - x1)
long_side = max(lh, lw)
h, w = mask.shape[:2]
if long_side < 1:
return 0
x_lst = np.round(np.linspace(x1, x2, long_side)).astype(np.int32)
y_lst = np.round(np.linspace(y1, y2, long_side)).astype(np.int32)
valid_pnts = (x_lst < w) & (y_lst < h)
if valid_pnts.sum() == 0:
return 0
score = mask[(y_lst[valid_pnts], x_lst[valid_pnts])].sum()
if divide_long_side:
score /= long_side
return score
# pass
def assign_mask_to_armature(mask, pos_list, score_list, score_thr=0.12, selected_armatures=None):
_valid_armatures = {
"handwear_0": (10, 8),
"handwear_1": (8, 6),
"topwear": (6, 5),
"handwear_2": (5, 7),
"handwear_3": (7, 9),
"legwear_1": (16, 14),
"legwear_2": (14, 12),
"bottomwear": (12, 11),
"legwear_3": (11, 13),
"legwear_4": (13, 15)
}
valid_armatures = {}
for k, vs in _valid_armatures.items():
is_valid = True
for v in vs:
if score_list[v] < score_thr:
is_valid = False
break
if is_valid:
valid_armatures[k] = vs
# armature_scores = {}
max_mask_score = 0
armature_assigned = None
assign_scores = {}
for k, vs in valid_armatures.items():
mask_score = mask_line_sample(mask, pos_list[vs[0]], pos_list[vs[1]], divide_long_side=True)
if mask_score > max_mask_score:
max_mask_score = mask_score
armature_assigned = k
assign_scores[k] = mask_score
armature_ids = None
if armature_assigned is not None:
armature_ids = valid_armatures[armature_assigned]
armature_assigned = armature_assigned.split('_')[0]
if selected_armatures is not None and armature_assigned not in selected_armatures:
return None, None, None
return armature_assigned, armature_ids, assign_scores
def armature_cc(mask_input, pos, armature_pairs, min_num_cc=-1):
retval, labels, stats, centroids = cv2.connectedComponentsWithStats(mask_input.astype(np.uint8), connectivity=8)
mask = None
if retval > min_num_cc:
# leg_pairs = [(16, 14), (14, 12), (11, 13), (13, 15)]
mask = np.zeros_like(labels, dtype=bool)
for l in range(1, retval):
mask_assigned = False
lmsk = l == labels
for p in armature_pairs:
if mask_line_sample(lmsk, pos[p[0]], pos[p[1]]) > 0:
mask_assigned = True
break
if mask_assigned:
mask = mask | lmsk
return mask
def taglist_has_keywords(tag_list, keywords):
if isinstance(keywords, str):
keywords = [keywords]
for tag in tag_list:
for k in keywords:
if k in tag:
return True
return False
TARGET_FRAME_SIZE = 1024
@cli.command('parse_render_body_samples')
@click.option('--exec_list')
@click.option('--bg_list')
@click.option('--save_dir')
@click.option('--rank_to_worldsize', default='', type=str)
def parse_render_body_samples(exec_list, bg_list, save_dir, rank_to_worldsize):
from live2d.scrap_model import animal_ear_detected, Drawable
from utils.cv import fgbg_hist_matching, quantize_image, random_crop, rle2mask, mask2rle, img_alpha_blending, resize_short_side_to, batch_save_masks, batch_load_masks
from utils.torch_utils import seed_everything
def _compose_body_samples(lmodel: Live2DScrapModel):
'''
some augmentation can be done here
'''
part_mask_list = []
body_final = lmodel.compose_bodypart_drawables(VALID_BODY_PARTS_V1)
for tag in VALID_BODY_PARTS_V1:
m = lmodel.compose_bodypart_drawables(tag, mask_only=True, final_visible_mask=True).astype(np.uint8)
# save_tmp_img(m, mask2img=True)
part_mask_list.append(m)
return True, part_mask_list, body_final
os.makedirs(save_dir, exist_ok=True)
seed_everything(42)
hist_match_prob = 0.2
quantize_prob = 0.25
color_correction_sampler = NameSampler({'hist_match': hist_match_prob, 'quantize': quantize_prob})
exec_list = load_exec_list(exec_list, rank_to_worldsize=rank_to_worldsize)
bg_list = load_exec_list(bg_list)
tagcluster_bodypart = json2dict('workspace/datasets/tagcluster_bodypart.json')
tag2generaltag = {}
for general_tag, tlist in tagcluster_bodypart.items():
for t in tlist:
if t in tag2generaltag and tag2generaltag[t] != general_tag:
print(f'conflict tag def: {t} - {general_tag}, ' + tag2generaltag[t])
tag2generaltag[t] = general_tag
for ii, p in enumerate(tqdm(exec_list[0:])):
try:
instance_mask, crop_xyxy, score = load_detected_character(p)
if instance_mask is None:
print(f'skip {p}, no character instance detected')
continue
lmodel = Live2DScrapModel(p, crop_xyxy=crop_xyxy, pad_to_square=False)
model_dir = lmodel.directory
if len(lmodel.facedet) == 0:
print(f'skip {model_dir}, no face detected')
pos_estim = load_pos_estimation(model_dir)
pos = pos_estim['pos']
pos_scores = pos_estim['scores']
if pos is None:
print(f'skip {model_dir}, no pos detected')
continue
pos = np.round(pos[:, ::-1]).astype(np.int32)
feet_valid = pos_scores[15] > 0.12 or pos_scores[16] > 0.12
tag_loaded = lmodel.load_tag_stats()
if not tag_loaded:
print(f'skip {model_dir}, no valid tag stats')
continue
tags = json2dict(osp.join(model_dir, 'general_tags.json'))
general_tags = set([tag2generaltag[k] for k in tags if k in tag2generaltag])
has_animal_ear = animal_ear_detected(tags)
face_parsing_loaded = lmodel.load_face_parsing()
if not face_parsing_loaded:
print(f'skip {face_parsing_loaded}, no valid faceparsing result')
continue
fx1, fy1, fx2, fy2 = lmodel.facedet[0]['bbox'][:4]
lmodel.init_drawable_visible_map()
frame_h, frame_w = lmodel.final.shape[:2]
wrist_left, wrist_right = pos[9], pos[10]
lang_sam_masks = None
lang_samp = osp.join(model_dir, 'langsam_masks.json')
if osp.exists(lang_samp):
lang_sam_masks = json2dict(lang_samp)['instances']
top_wear_assigned = hand_msk_assigned = mouth_msk_assigned = nose_msk_assigned = neck_msk_assigned = False
if lang_sam_masks is not None:
if has_animal_ear:
ear_msk = [m for m in lang_sam_masks['ears']['masks']]
if len(ear_msk) > 0:
_, ear_msk_assigned = try_assign_sam_mask(
lmodel,
[rle2mask(m) for m in ear_msk], body_part_tag='ears',
exclude_mask=lmodel.compose_face_drawables(face_part_ids=[7, 8], mask_only=True, final_visible_mask=True), mask_ioa_thr=0.3,
valid_face_ids={17, 18}, y2_max=fy2
)
_, hand_msk_assigned = try_assign_sam_mask(
lmodel,
[rle2mask(m) for m in lang_sam_masks['hand']['masks']],
check_points=[wrist_left, wrist_right], body_part_tag='handwear', mask_ioa_thr=0.1, valid_face_ids={16, 17}
)
ankle_kpts = [pos[k] for k in [15, 16] if pos_scores[k] > 0.12]
if len(ankle_kpts) > 0:
_, feet_msk_assigned = try_assign_sam_mask(
lmodel,
[rle2mask(m) for m in lang_sam_masks['feet']['masks']],
check_points=ankle_kpts, body_part_tag='footwear'
)
# _, shoes_msk_assigned = try_assign_sam_mask(
# lmodel,
# [rle2mask(m) for m in lang_sam_masks['shoes']['masks']],
# check_points=[ankle_left, ankle_right], body_part_tag='footwear'
# )
# save_tmp_img(visualize_segs([rle2mask(m) for m in lang_sam_masks['shoes']['masks']], lmodel.final[..., :3]))
# save_tmp_img(visualize_segs([rle2mask(m) for m in lang_sam_masks['feet']['masks']], lmodel.final[..., :3]))
# feet_msk_assigned = feet_msk_assigned or shoes_msk_assigned
top_wear_mask = None
top_wear_masks = lang_sam_masks['jacket']['masks'] + lang_sam_masks['dress']['masks']
top_wear_scores = lang_sam_masks['jacket']['scores'] + lang_sam_masks['dress']['scores']
if len(top_wear_scores) > 0:
midx = np.argmax(np.array(top_wear_scores))
top_wear_mask = rle2mask(top_wear_masks[midx])
if top_wear_mask is not None:
top_wear_masks = [top_wear_mask]
else:
top_wear_masks = []
# top_wear_masks += [rle2mask(m) for m in lang_sam_masks['shirt']['masks']]
if top_wear_masks is not None:
_, top_wear_assigned = try_assign_sam_mask(
lmodel,
top_wear_masks,
body_part_tag='topwear',
valid_face_ids={16}
)
_, leg_msk_assigned = try_assign_sam_mask(
lmodel,
[rle2mask(m) for m in lang_sam_masks['leg']['masks']],
body_part_tag='legwear'
)
# save_tmp_img(visualize_segs([rle2mask(m) for m in lang_sam_masks['hand']['masks']], lmodel.final[..., :3]))
_, hair_msk_assigned = try_assign_sam_mask(
lmodel,
[rle2mask(m) for m in lang_sam_masks['hair']['masks']],
body_part_tag='hair',
valid_face_ids={17}
)
_, face_msk_assigned = try_assign_sam_mask(
lmodel,
[rle2mask(m) for m in lang_sam_masks['face']['masks']],
body_part_tag='face',
valid_face_ids={1}
)
_, neck_msk_assigned = try_assign_sam_mask(
lmodel,
[rle2mask(m) for m in lang_sam_masks['face']['masks']],
body_part_tag='neck',
valid_face_ids={14, 16,}, mask_ioa_thr=0.3
)
_, nose_msk_assigned = try_assign_sam_mask(
lmodel,
[rle2mask(m) for m in lang_sam_masks['nose']['masks']],
body_part_tag='nose',
valid_face_ids={10, 11, 1}, mask_ioa_thr=0.3,
bbox_constraint=lmodel.facedet[0]['bbox'][:4]
)
_, mouth_msk_assigned = try_assign_sam_mask(
lmodel,
[rle2mask(m) for m in lang_sam_masks['mouth']['masks']],
body_part_tag='mouth',
valid_face_ids={10, 11, 1}, mask_ioa_thr=0.3,
bbox_constraint=lmodel.facedet[0]['bbox'][:4]
)
left_out_drawables: list[Drawable] = []
for d in lmodel.drawables:
if d.final_visible_area < 1:
continue
if d.face_part_id in {2, 3, 4, 5}: # eyes, eyebrows
d.body_part_tag = 'eyes'
if d.face_part_id in {6}: # glasses
d.body_part_tag = 'eyewear'
if d.face_part_id == 17: # hair
if d.body_part_tag is None:
d.body_part_tag = 'hair'
elif d.body_part_tag == 'ears':
d.face_part_id = 7
if d.face_part_id == 18: # hat
if d.body_part_tag is None:
d.body_part_tag = 'headwear'
elif d.body_part_tag == 'ears':
d.face_part_id = 7
if d.face_part_id in {7, 8}: # ears
d.body_part_tag = 'ears'
if d.face_part_id == 16: # cloth
if d.body_part_tag is None:
d.body_part_tag = 'topwear'
elif d.body_part_tag == 'neck':
d.face_part_id = 14
if d.face_part_id == 14: # neck
if not neck_msk_assigned:
d.body_part_tag = 'neck'
elif d.body_part_tag == 'topwear':
d.face_part_id = 16
if not mouth_msk_assigned:
if d.face_part_id == 11:
d.body_part_tag = 'mouth'
if not nose_msk_assigned:
if d.face_part_id == 10:
d.body_part_tag = 'nose'
if d.body_part_tag is None and d.tag_stats is not None:
left_out_drawables.append(d)
GENERAL_TAGS_FOR_LEFTOUT = [
'headwear', 'eyewear', 'earwear', 'beard', 'neckwear',
'skin', 'topwear', 'handwear',
'bottomwear', 'legwear', 'footwear',
'tail', 'wings'
]
def tagstats_to_general_tagstats(tagstats, valid_gtags=None):
general_tagstats = {}
if valid_gtags is not None:
valid_gtags = set(valid_gtags)
for t, stats in tagstats.items():
if t not in tag2generaltag:
continue
gt = tag2generaltag[t]
if valid_gtags is not None and gt not in valid_gtags:
continue
if gt in general_tagstats:
if general_tagstats[gt]['avg_score'] < stats['avg_score']:
general_tagstats[gt] = stats
else:
general_tagstats[gt] = stats
return general_tagstats
for d in left_out_drawables:
tagstats = tagstats_to_general_tagstats(d.tag_stats, valid_gtags=GENERAL_TAGS_FOR_LEFTOUT)
sorted_items = sorted(tagstats.items(), key=lambda item: item[1]['avg_score'], reverse=True)
tagstats = dict(sorted_items)
if 'handwear' in tagstats:
if not mask_cover_pos(d.final_visible_mask, pos, [5, 6, 7, 8, 9, 10], xshift=-d.x, yshift=-d.y):
if mask_cover_pos(d.final_visible_mask, pos, [11, 12], xshift=-d.x, yshift=-d.y):
tagstats.pop('handwear')
tagstats_lst = list(tagstats.keys())
if len(tagstats) > 1 and tagstats[tagstats_lst[0]]['avg_score'] > 0.1:
d.body_part_tag = tagstats_lst[0]
if d.body_part_tag is None:
assigned_armature, _, _ = assign_mask_to_armature(d.get_full_mask(final_visible_mask=True), pos, pos_scores)
if assigned_armature is not None:
d.body_part_tag = assigned_armature
for d in lmodel.drawables:
if d.body_part_tag == 'topwear':
tagstats = tagstats_to_general_tagstats(d.tag_stats, valid_gtags=['topwear', 'bottomwear', 'legwear', 'handwear', 'hair'])
max_score_tag = max(list(tagstats.keys()), key=lambda x: tagstats[x]['avg_score'])
# d.body_part_tag = max_score_tag
if max_score_tag == 'handwear':
coords = np.where(d.final_visible_mask)
cx, cy = np.mean(coords[1]), np.mean(coords[0])
cy += d.y
if cy < np.mean(pos[[11, 12, 13, 14]], axis=0)[1]:
d.body_part_tag = max_score_tag
elif max_score_tag in {'legwear', 'bottomwear'}:
coords = np.where(d.final_visible_mask)
x1, y1, x2, y2 = np.min(coords[1]), np.min(coords[0]), np.max(coords[1]), np.max(coords[0])
cx, cy = (x1 + x2) / 2, (y1 + y2) / 2
cy += d.y
if cy > np.mean(pos[[11, 12]], axis=0)[1]:
d.body_part_tag = max_score_tag
else:
d.body_part_tag = max_score_tag
head_y = np.mean(pos[[0, 1, 2, 3, 4]], axis=0)[1]
shoulder_mid = np.mean(pos[[5, 6]], axis=0)[0]
shoulder_y = np.mean(pos[[5, 6], 1], axis=0)
shoulder_x1, shoulder_x2 = np.min(pos[[5, 6]], axis=0)[0], np.max(pos[[5, 6]], axis=0)[0]
head_order = min([d.draw_order for d in lmodel.drawables if d.body_part_tag == 'face'])
hip_y = np.mean(pos[[11, 12]], axis=0)[1]
hip_x1, hip_x2 = pos[11][0], pos[11][1]
if hip_x1 > hip_x2:
hip_x1, hip_x2 = hip_x2, hip_x1
hip_mid = (hip_x1 + hip_x2) / 2
mask = lmodel.compose_bodypart_drawables('handwear', mask_only=True, final_visible_mask=True)
# save_tmp_img(mask, mask2img=True)
hair_drawabls = lmodel.get_body_part_drawables('hair')
topwear_drawabls = lmodel.get_body_part_drawables('topwear')
# topwear_order = None
# if len(topwear_drawabls) > 0:
# topwear_order = max([d.draw_order for d in topwear_drawabls])
for d in hair_drawabls + topwear_drawabls:
x1, y1, x2, y2 = d.visible_xyxy
dh = y2 - y1
if (x1 - shoulder_mid) * (x2 - shoulder_mid) < 0:
continue
if min(shoulder_x2, x2) - max(shoulder_x1, x1) > 0.3 * (shoulder_x2 - shoulder_x1):
continue
# if d.draw_order > topwear_order:
# continue
d_midx = (x1 + x2) / 2
if d_midx > shoulder_x1 and d_midx < shoulder_x2 and d.body_part_tag == 'topwear':
continue
if d.y < head_y:
continue
assigned_armature, armature_ids, assign_scores = assign_mask_to_armature(
d.get_full_mask(final_visible_mask=True), pos, pos_scores, selected_armatures={'handwear'})
if assigned_armature is not None:
armature_scores = []
if 8 in armature_ids:
armature_h = np.max(pos[[6, 8, 10], 1]) - np.min(pos[[6, 8, 10], 1])
for h in ['handwear_0', 'handwear_1']:
if h in assign_scores:
armature_scores.append(assign_scores[h])
else:
armature_h = np.max(pos[[5, 7, 9], 1]) - np.min(pos[[5, 7, 9], 1])
for h in ['handwear_2', 'handwear_3']:
if h in assign_scores:
armature_scores.append(assign_scores[h])
armature_length = np.linalg.norm(pos[armature_ids[0]] - pos[armature_ids[1]])
if not (len(armature_scores) > 1 and np.mean(armature_scores) > 0.8) and dh > armature_length * 1.5 and d.body_part_tag == 'topwear':
continue
# print(int(d_midx), pos[5][1])
# save_tmp_img(cv2.circle(lmodel.final[..., :3].copy(), (int(d_midx), pos[5][1]), 15, (255, 255, 0), thickness=-1))
d.body_part_tag = assigned_armature
for d in lmodel.drawables:
if d.body_part_tag == 'handwear':
if d.y < head_y and d.draw_order < head_order:
d.body_part_tag = 'hair'
if d.body_part_tag == 'legwear':
x1, y1, x2, y2 = d.visible_xyxy
if y2 < hip_y:
d.body_part_tag = 'topwear'
if d.body_part_tag in {'neck', 'neckwear', 'headwear'}:
x1, y1, x2, y2 = d.visible_xyxy
if y1 > hip_y:
d.body_part_tag = 'topwear'
legmask = armature_cc(
lmodel.compose_bodypart_drawables('legwear', mask_only=True, final_visible_mask=True),
pos, [(16, 14), (14, 12), (11, 13), (13, 15)], min_num_cc=3
)
dr_x1, dr_y1, dr_x2, dr_y2 = mask_xyxy(lmodel.compose_bodypart_drawables(['topwear', 'bottomwear'], mask_only=True, final_visible_mask=True))
if legmask is not None:
for d in lmodel.drawables:
x1, y1, x2, y2 = d.xyxy
if d.body_part_tag != 'legwear':
continue
if not np.any(legmask[y1: y2, x1: x2] & d.final_visible_mask):
d.body_part_tag = 'topwear'
bottommask = armature_cc(
lmodel.compose_bodypart_drawables('bottomwear', mask_only=True, final_visible_mask=True),
pos, [(11, 12), (14, 12), (11, 13)]
)
if not np.any(bottommask):
for d in lmodel.drawables:
if d.body_part_tag == 'bottomwear':
d.body_part_tag = 'topwear'
if len(lmodel.get_body_part_drawables('bottomwear')) == 0:
for d in lmodel.drawables:
if not d.body_part_tag == 'topwear':
continue
x1, y1, x2, y2 = d.visible_xyxy
ix1, ix2 = max(x1, hip_x1), min(x2, hip_x2)
if (ix2 - ix1) / (hip_x2 - hip_x1) < 0.9:
continue
my = (y1 + y2) / 2
if my > hip_y or (hip_y - my) < (my - shoulder_y) / 5 and y1 > np.min(pos[[5, 6], 1]):
d.body_part_tag = 'bottomwear'
bt_x1, bt_y1, bt_x2, bt_y2 = mask_xyxy(lmodel.compose_bodypart_drawables(['bottomwear'], mask_only=True, final_visible_mask=True))
neck_drawables = lmodel.get_body_part_drawables('neck')
if len(neck_drawables) > 0:
neck_order = min(d.draw_order for d in neck_drawables)
neck_mask = lmodel.compose_bodypart_drawables('neck', mask_only=True, final_visible_mask=False)
# save_tmp_img(neck_mask, mask2img=True)
neck_mask2 = lmodel.compose_bodypart_drawables('neck', mask_only=True, final_visible_mask=True)
n2x1, n2y1, n2x2, n2y2 = mask_xyxy(neck_mask2)
neck_mask[:n2y1] = 0
nx1, ny1, nx2, ny2 = mask_xyxy(neck_mask)
neck_mask = neck_mask[ny1: ny2, nx1: nx2]
nh = ny2 - ny1
# dr_x1, dr_y1, dr_x2, dr_y2 = mask_xyxy(lmodel.compose_bodypart_drawables(['topwear', 'bottomwear'], mask_only=True, final_visible_mask=True))
lg_x1, lg_y1, lg_x2, lg_y2 = mask_xyxy(lmodel.compose_bodypart_drawables(['legwear'], mask_only=True, final_visible_mask=True))
no_leg = lg_x1 == 0 and lg_x2 == 0 and lg_y1 == 0 and lg_y2 == 0
feet_drawable_exists = False
for d in lmodel.drawables:
if d.body_part_tag is None:
continue
x1, y1, x2, y2 = d.visible_xyxy
my = (y1 + y2) / 2
dh = y2 - y1
if d.body_part_tag in {'footwear', 'wings'}:
if x1 > dr_x1 and y1 > dr_y1 and x2 < dr_x2 and y2 < dr_y2:
d.body_part_tag = 'topwear'
continue
elif d.body_part_tag == 'footwear':
leg_len = (13, 15) if (x1 + x2) / 2 < hip_mid else (16, 14)
leg_len = np.linalg.norm(pos[leg_len[0]] - pos[leg_len[1]])
if np.abs(y1 - y2) > leg_len * 1.0:
d.body_part_tag = 'legwear'
elif no_leg:
d.body_part_tag = 'legwear'
elif np.abs(y1 - y2) > 2 * (lg_y2 - lg_y1):
d.body_part_tag = 'legwear'
elif d.body_part_tag in {'hair'}:
if y1 > fy2 and (x1 - shoulder_mid) * (x2 - shoulder_mid) < 0 and x1 > dr_x1 and x2 < dr_x2 and y1 > dr_y1 and y2 < dr_y2:
d.body_part_tag = 'topwear'
elif len(neck_drawables) > 0 and neck_order < d.draw_order and y1 > ny1 - nh:
mask_i = d.bitwise_and(neck_mask, [nx1, ny1, nx2, ny2], final_vis_mask=True)
if np.any(mask_i):
d.body_part_tag = 'topwear'
# d.body_part_tag =
elif d.body_part_tag == 'neckwear':
if (hip_y - my) < (my - shoulder_y):
d.body_part_tag = 'topwear'
elif d.body_part_tag == 'topwear':
if bt_y1 > 0 and y1 > bt_y1 and y2 < bt_y2 and x1 > bt_x1 and x2 < bt_x2:
d.body_part_tag = 'bottomwear'
if d.body_part_tag == 'footwear':
if feet_valid:
feet_drawable_exists = True
else:
d.body_part_tag = 'legwear'
feet_mask_valid = feet_drawable_exists == feet_valid
# print(feet_drawable_exists, feet_valid, feet_mask_valid)
metadata = {'tag_valid': {k: True for k in VALID_BODY_PARTS_V1}}
metadata['tag_valid']['footwear'] = feet_mask_valid
lmodel.save_body_parsing(metadata=metadata)
tagged_drawables = [d for d in lmodel.drawables if d.body_part_tag in VALID_BODY_PARTS_V1]
init_drawable_visible_map(tagged_drawables)
is_valid, masks, final_img = _compose_body_samples(lmodel,)
reference_img = lmodel.final[..., :3]
# from utils.visualize import visualize_pos_keypoints
# savep = None
# savep = osp.join('workspace/segs', osp.basename(model_dir) + '.png')
# imwrite(savep, reference_img)
# savep = osp.join('workspace/segs', osp.basename(model_dir) + '_segs.png')
# imwrite(savep, visualize_segs_with_labels(masks, final_img, tag_list=VALID_BODY_PARTS_V1, image_weight=0.0, draw_legend=False))
# savep = osp.join('workspace/cases/output', osp.basename(model_dir) + '.png')
# reference_img = np.array(visualize_pos_keypoints(reference_img, keypoints=pos[..., ::-1]))
# save_tmp_img(visualize_segs_with_labels(masks, final_img, tag_list=VALID_BODY_PARTS_V1, image_weight=0.0, reference_img=reference_img))
foot_msk_idx = VALID_BODY_PARTS_V1.index('footwear')
leg_msk_idx = VALID_BODY_PARTS_V1.index('legwear')
masks[leg_msk_idx] = masks[leg_msk_idx] | masks[foot_msk_idx]
bgp = random.choice(bg_list)
fh, fw = final_img.shape[:2]
bg = imread(bgp)
fsize = max(fh, fw)
target_bg_size = max(fsize, TARGET_FRAME_SIZE)
fsze_max = int(round(fsize * 1.6))
if fsze_max != target_bg_size:
target_bg_size = random.randint(min(fsze_max, target_bg_size), max(fsze_max, target_bg_size))
bg = resize_short_side_to(bg, target_bg_size)
bg = random_crop(imread(bgp), (target_bg_size, target_bg_size))
if fh != target_bg_size or fw != target_bg_size:
px = py = 0
if fh != target_bg_size:
py = random.randint(0, target_bg_size - fh)
if fw != target_bg_size:
px = random.randint(0, target_bg_size - fw)
blank_final = np.zeros((target_bg_size, target_bg_size, 4), np.uint8)
blank_final[py: py + fh, px: px + fw] = final_img
final_img = blank_final
for mi, m in enumerate(masks):
blank = np.zeros((target_bg_size, target_bg_size), bool)
blank[py: py + fh, px: px + fw] = m
masks[mi] = blank
fh, fw = final_img.shape[:2]
color_correct = color_correction_sampler.sample()
if color_correct == 'hist_match':
fgbg_hist_matching([final_img], bg)
face_wbg = img_alpha_blending([bg, final_img])
if color_correct == 'quantize':
mask = final_img[..., -1] > 35
# cv2.imshow("mask", mask)
face_wbg[..., :3] = quantize_image(face_wbg[..., :3], random.choice([12, 16, 32]), 'kmeans', mask=mask)[0]
d = osp.abspath(model_dir).replace('\\', '/').rstrip('/').replace('.', '_DOT_')
d1 = d.split('/')[-1]
d2 = d.split('/')[-3]
savename = d2 + '____' + d1
savep = osp.join(save_dir, savename)
# save_tmp_img(face_wbg)
imwrite(savep, face_wbg, quality=97, ext='.jpg')
mask_meta_list = [{} for _ in range(len(VALID_BODY_PARTS_V1))] # dont use [{}] * len
mask_meta_list[foot_msk_idx]['is_valid'] = feet_mask_valid
batch_save_masks(masks, savep + '.json', mask_meta_list=mask_meta_list)
except Exception as e:
# raise
print(f'Failed to process {p}: {e}')
@cli.command('parse_render_body_samples_wsegs')
@click.option('--exec_list')
@click.option('--bg_list')
@click.option('--mask_name')
@click.option('--save_dir', default='')
@click.option('--rank_to_worldsize', default='', type=str)
def parse_render_body_samples_wsegs(exec_list, bg_list, mask_name, save_dir, rank_to_worldsize):
from live2d.scrap_model import animal_ear_detected, Drawable
from utils.cv import fgbg_hist_matching, quantize_image, random_crop, rle2mask, mask2rle, img_alpha_blending, resize_short_side_to, batch_save_masks, batch_load_masks
from utils.torch_utils import seed_everything
def _compose_body_samples(lmodel: Live2DScrapModel):
'''
some augmentation can be done here
'''
part_mask_list = []
body_final = lmodel.compose_bodypart_drawables(VALID_BODY_PARTS_V1)
for tag in VALID_BODY_PARTS_V1:
m = lmodel.compose_bodypart_drawables(tag, mask_only=True, final_visible_mask=True).astype(np.uint8)
# save_tmp_img(m, mask2img=True)
part_mask_list.append(m)
return True, part_mask_list, body_final
seed_everything(42)
hist_match_prob = 0.2
quantize_prob = 0.25
color_correction_sampler = NameSampler({'hist_match': hist_match_prob, 'quantize': quantize_prob})
exec_list = load_exec_list(exec_list, rank_to_worldsize=rank_to_worldsize)
bg_list = load_exec_list(bg_list)
tagcluster_bodypart = json2dict('workspace/datasets/tagcluster_bodypart.json')
tag2generaltag = {}
for general_tag, tlist in tagcluster_bodypart.items():
for t in tlist:
if t in tag2generaltag and tag2generaltag[t] != general_tag:
print(f'conflict tag def: {t} - {general_tag}, ' + tag2generaltag[t])
tag2generaltag[t] = general_tag
if save_dir != '':
os.makedirs(save_dir, exist_ok=True)
render_sample = save_dir != ''
for ii, p in enumerate(tqdm(exec_list[0:])):
try:
instance_mask, crop_xyxy, score = load_detected_character(p)
if instance_mask is None:
print(f'skip {p}, no character instance detected')
continue
lmodel = Live2DScrapModel(p, crop_xyxy=crop_xyxy, pad_to_square=False)
model_dir = lmodel.directory
if len(lmodel.facedet) == 0:
print(f'skip {model_dir}, no face detected')
pos_estim = load_pos_estimation(model_dir)
pos = pos_estim['pos']
pos_scores = pos_estim['scores']
if pos is None:
print(f'skip {model_dir}, no pos detected')
continue
pos = np.round(pos[:, ::-1]).astype(np.int32)
tag_loaded = lmodel.load_tag_stats()
if not tag_loaded:
print(f'skip {model_dir}, no valid tag stats')
continue
tags = json2dict(osp.join(model_dir, 'general_tags.json'))
general_tags = set([tag2generaltag[k] for k in tags if k in tag2generaltag])
has_animal_ear = animal_ear_detected(tags)
face_parsing_loaded = lmodel.load_face_parsing()
if not face_parsing_loaded:
print(f'skip {model_dir}, no face_parsing')
continue
bodyparsing_loaded = lmodel.load_body_parsing()
if not bodyparsing_loaded:
print(f'skip {model_dir}, no bodyparsing_loaded')
continue
head_y = np.mean(pos[[0, 1, 2, 3, 4]], axis=0)[1]
shoulder_mid = np.mean(pos[[5, 6]], axis=0)[0]
shoulder_y = np.mean(pos[[5, 6], 1], axis=0)
shoulder_x1, shoulder_x2 = np.min(pos[[5, 6]], axis=0)[0], np.max(pos[[5, 6]], axis=0)[0]
head_order = min([d.draw_order for d in lmodel.drawables if d.body_part_tag == 'face'])
hip_y = np.mean(pos[[11, 12]], axis=0)[1]
hip_x1, hip_x2 = pos[11][0], pos[11][1]
belly_y = (shoulder_y + hip_y) / 2
metadata = lmodel._body_parsing['metadata']
feet_mask_valid = metadata['tag_valid']['footwear']
masks_ann = json2dict(osp.join(model_dir, mask_name + '.json'))
sam_masks = [rle2mask(m, to_bool=True) for m in masks_ann]
tagged_drawables = [d for d in lmodel.drawables if d.body_part_tag in VALID_BODY_PARTS_V1]
init_drawable_visible_map(tagged_drawables)
for tg in tagged_drawables:
if tg.body_part_tag in {'eyewear', 'nose', 'mouth', 'eyes', 'face', 'tail'}:
continue
ori_tag = tg.body_part_tag
if feet_mask_valid and ori_tag == 'footwear':
continue
score_list = []
for m in sam_masks:
area, u_area, i_area = tg.mask_union_intersection(m, final_vis_mask=True)
if i_area is None:
i_area = -1
score = i_area / tg.final_visible_area
score_list.append(score)
best_match = np.argmax(np.array(score_list))
best_match = VALID_BODY_PARTS_V1[best_match]
tg.body_part_tag = best_match
if tg.body_part_tag == 'legwear' and score_list[VALID_BODY_PARTS_V1.index('footwear') > 0.5]:
tg.body_part_tag = 'footwear'
if ori_tag in {'bottomwear', 'legwear', 'footwear'} and tg.body_part_tag in {'handwear', 'hair'}:
tg.body_part_tag = ori_tag
for tg in tagged_drawables:
x1, y1, x2, y2 = tg.xyxy
if tg.body_part_tag in {'headwear', 'neck', 'neckwear'} and (y1 + y2) / 2 > belly_y:
tg.body_part_tag = 'topwear'
lmodel.save_body_parsing(metadata=metadata, save_name='parsinglog_' + mask_name)
if not render_sample:
continue
is_valid, masks, final_img = _compose_body_samples(lmodel,)
reference_img = lmodel.final[..., :3]
from utils.visualize import visualize_pos_keypoints
savep = None
# savep = osp.join('workspace/segs', osp.basename(model_dir) + '.png')
# imwrite(savep, reference_img)
# savep = osp.join('workspace/segs', osp.basename(model_dir) + '_segs.png')
# imwrite(savep, visualize_segs_with_labels(masks, final_img, tag_list=VALID_BODY_PARTS_V1, image_weight=0.0, draw_legend=True, reference_img=reference_img))
# savep = osp.join('workspace/cases24k', osp.basename(model_dir) + '.png')
# reference_img = np.array(visualize_pos_keypoints(reference_img, keypoints=pos[..., ::-1]))
# save_tmp_img(visualize_segs_with_labels(masks, final_img, tag_list=VALID_BODY_PARTS_V1, image_weight=0.0, reference_img=reference_img), savep=savep)
# continue
foot_msk_idx = VALID_BODY_PARTS_V1.index('footwear')
leg_msk_idx = VALID_BODY_PARTS_V1.index('legwear')
masks[leg_msk_idx] = masks[leg_msk_idx] | masks[foot_msk_idx]
bgp = random.choice(bg_list)
fh, fw = final_img.shape[:2]
bg = imread(bgp)
fsize = max(fh, fw)
target_bg_size = max(fsize, TARGET_FRAME_SIZE)
fsze_max = int(round(fsize * 1.6))
if fsze_max != target_bg_size:
target_bg_size = random.randint(min(fsze_max, target_bg_size), max(fsze_max, target_bg_size))
bg = resize_short_side_to(bg, target_bg_size)
bg = random_crop(imread(bgp), (target_bg_size, target_bg_size))
if fh != target_bg_size or fw != target_bg_size:
px = py = 0
if fh != target_bg_size:
py = random.randint(0, target_bg_size - fh)
if fw != target_bg_size:
px = random.randint(0, target_bg_size - fw)
blank_final = np.zeros((target_bg_size, target_bg_size, 4), np.uint8)
blank_final[py: py + fh, px: px + fw] = final_img
final_img = blank_final
for mi, m in enumerate(masks):
blank = np.zeros((target_bg_size, target_bg_size), bool)
blank[py: py + fh, px: px + fw] = m
masks[mi] = blank
fh, fw = final_img.shape[:2]
color_correct = color_correction_sampler.sample()
if color_correct == 'hist_match':
fgbg_hist_matching([final_img], bg)
face_wbg = img_alpha_blending([bg, final_img])
if color_correct == 'quantize':
mask = final_img[..., -1] > 35
# cv2.imshow("mask", mask)
face_wbg[..., :3] = quantize_image(face_wbg[..., :3], random.choice([12, 16, 32]), 'kmeans', mask=mask)[0]
d = osp.abspath(model_dir).replace('\\', '/').rstrip('/').replace('.', '_DOT_')
d1 = d.split('/')[-1]
d2 = d.split('/')[-3]
savename = d2 + '____' + d1
savep = osp.join(save_dir, savename)
# save_tmp_img(face_wbg)
imwrite(savep, face_wbg, quality=97, ext='.jpg')
mask_meta_list = [{} for _ in range(len(VALID_BODY_PARTS_V1))] # dont use [{}] * len
mask_meta_list[foot_msk_idx]['is_valid'] = feet_mask_valid
batch_save_masks(masks, savep + '.json', mask_meta_list=mask_meta_list)
except Exception as e:
# raise
print(f'Failed to process {p}: {e}')
@cli.command('render_face_samples')
@click.option('--exec_list')
@click.option('--bg_list')
@click.option('--save_dir')
@click.option('--rank_to_worldsize', default='', type=str)
def render_face_samples(exec_list, bg_list, save_dir, rank_to_worldsize):
TARGET_FRAME_SIZE = 2048
from utils.cv import fgbg_hist_matching, quantize_image, random_crop, img_bbox, img_alpha_blending, resize_short_side_to, batch_save_masks, batch_load_masks
from utils.torch_utils import seed_everything
from utils.visualize import FACE_LABEL2NAME
def _compose_face_samples(lmodel: Live2DScrapModel):
'''
todo: save complete part
'''
face_xyxy = lmodel.face_seg_xyxy
face_h, face_w = face_xyxy[3] - face_xyxy[1], face_xyxy[2] - face_xyxy[0]
all_face_labels = list(FACE_LABEL2NAME.keys())
face_final = lmodel.compose_face_drawables(list(FACE_LABEL2NAME.keys()), xyxy=face_xyxy)
# save_tmp_img(face_final, 'local_tmp.png')
part_mask_list = []
# segmap = np.zeros((face_h, face_w), dtype=np.int32)
alphas = np.zeros((face_h, face_w), dtype=np.int32)
for ii in range(1, len(all_face_labels)):
m = lmodel.compose_face_drawables(ii, mask_only=True, xyxy=face_xyxy, final_visible_mask=True).astype(np.uint8)
# save_tmp_img(m, mask2img=True)
part_mask_list.append(m)
mask_bg = np.bitwise_not(np.bitwise_or.reduce(np.stack(part_mask_list).astype(bool), axis=0))
part_mask_list.insert(0, mask_bg.astype(np.uint8))
nose_detected, mouth_detected = lmodel.face_part_detected([10, 11])
tp = osp.join(lmodel.directory, 'faceseg_nosemouth.json.gz')
if osp.exists(tp) and (not nose_detected or not mouth_detected):
nose_mouth = batch_load_masks(tp)
if not nose_detected:
part_mask_list[10] = nose_mouth[0]
part_mask_list[1][np.where(nose_mouth[0] > 0)] = 0
if not mouth_detected:
part_mask_list[11] = nose_mouth[1]
part_mask_list[1][np.where(nose_mouth[1] > 0)] = 0
bx, by, bw, bh = cv2.boundingRect(cv2.findNonZero(part_mask_list[0].astype(np.uint8)))
by2 = by + bh
bx2 = bw + bx
# DONT DELETE THESE!!!!
# depth_lower = 100000
# depth_upper = -1
# for d_id, drawable in enumerate(lmodel.drawables):
# if drawable.area < 1 or not drawable.face_part_id == 1:
# continue
# dx, dy, dw, dh = drawable.get_bbox(xyxy=face_xyxy)
# dx2 = dx + dw
# dy2 = dy + dh
# # check if hair drawable is actually background
# if drawable.face_part_id == 17:
# if drawable.face_part_stats['ioa'][0] > 0.7 and drawable.face_part_stats['ioa'][17] < 0.3:
# drawable.face_part_id = None
# if drawable.face_part_id == 1 and dw / bw > 0.7 and dh > bw > 0.7:
# if drawable.draw_order < depth_lower:
# depth_lower = drawable.draw_order
# if drawable.draw_order > depth_upper:
# depth_upper = drawable.draw_order
# depth_buffer = np.zeros((face_h, face_w), dtype=np.uint8)
# base_depth = 1
# mask = np.zeros_like(depth_buffer, dtype=bool)
# valid_face_ids = set(range(1, 19))
# for d in lmodel.drawables:
# if d.area < 1 or d.face_part_id not in valid_face_ids:
# continue
# if np.any(d.bitwise_and(mask, face_xyxy)):
# base_depth += 1
# m = d.get_full_mask(xyxy=face_xyxy)
# mask |= m
# d.depth = base_depth
# depth_buffer[np.where(m)] = base_depth
# depth = (depth_buffer / np.max(depth_buffer) * 255).astype(np.uint8)
# save_tmp_img(depth)
# base_face_mask = compose_from_drawables([d for d in lmodel.drawables if \
# drawable.draw_order >= depth_lower and drawable.draw_order > depth_upper])
# for drawable in lmodel.drawables:
# if drawable.draw_order < depth_lower or drawable.draw_order > depth_upper:
# continue
# segmap = segmap.astype(np.uint8)
# lmodel.compose_face_drawables([4, 5], output_type='pil').save('local_tst.png')
# save_tmp_img(face_final)
# save_tmp_img(segmap == 1, mask2img=True)
# save_tmp_img(segmap == 4, mask2img=True)
return True, part_mask_list, face_final
os.makedirs(save_dir, exist_ok=True)
seed_everything(42)
hist_match_prob = 0.2
quantize_prob = 0.25
color_correction_sampler = NameSampler({'hist_match': hist_match_prob, 'quantize': quantize_prob})
if exec_list.endswith('.json'):
new_exec_list = []
exec_list = json2dict(exec_list)
for k, vs in exec_list.items():
for v in vs:
new_exec_list.append({v: k})
exec_list = new_exec_list
pass
exec_list = load_exec_list(exec_list, rank_to_worldsize=rank_to_worldsize)
bg_list = load_exec_list(bg_list)
VALID_FACE_SET = set(range(19))
for ii, p in enumerate(tqdm(exec_list[0:])):
try:
face_parsingp = None
if isinstance(p, dict):
for k, v in p.items():
p = k
face_parsingp = osp.join(v, 'face_parsing.json')
lmodel = Live2DScrapModel(p)
model_dir = lmodel.directory
if face_parsingp is None:
face_parsingp = osp.join(model_dir, 'face_parsing.json')
if not osp.exists(face_parsingp):
face_parsingp = '-'.join(model_dir.split('-')[:-1]) + '-4'
face_parsingp = osp.join(face_parsingp, 'face_parsing.json')
if not osp.exists(face_parsingp):
print(f"skip {p} due to face parsing result not found")
continue
lmodel.load_face_parsing(face_parsingp)
face_drawables = [d for d in lmodel.drawables if d.face_part_id in VALID_FACE_SET]
init_drawable_visible_map(face_drawables)
is_valid, labels, face_final = _compose_face_samples(lmodel,)
mask_list = labels
if not is_valid:
continue
# save_tmp_img(labels[0], mask2img=True)
bgp = random.choice(bg_list)
fh, fw = face_final.shape[:2]
bg = imread(bgp)
bgh, bgw = bg.shape[:2]
target_bg_size = min(bgh, bgw, TARGET_FRAME_SIZE)
fsize = max(fh, fw)
if fsize * 2 < target_bg_size:
target_bg_size = random.randint(fsize * 2, target_bg_size)
bg = resize_short_side_to(bg, target_bg_size)
bg = random_crop(imread(bgp), (fh, fw))
# save_tmp_img(bg)
color_correct = color_correction_sampler.sample()
if color_correct == 'hist_match':
fgbg_hist_matching([face_final], bg)
face_wbg = img_alpha_blending([bg, face_final])
if color_correct == 'quantize':
mask = face_final[..., -1] > 35
# cv2.imshow("mask", mask)
face_wbg[..., :3] = quantize_image(face_wbg[..., :3], random.choice([12, 16, 32]), 'kmeans', mask=mask)[0]
d = osp.abspath(model_dir).replace('\\', '/').rstrip('/').replace('.', '_DOT_')
d1 = d.split('/')[-1]
d2 = d.split('/')[-3]
savename = d2 + '____' + d1
savep = osp.join(save_dir, savename)
# save_tmp_img(face_wbg)
imwrite(savep, face_wbg, quality=97, ext='.jpg')
batch_save_masks(mask_list, savep + '.json', compress='gzip')
# print(f'finished {savep}')
except Exception as e:
# raise
print(f'Failed to process {p}: {e}')
@cli.command('dump_clean_lst')
@click.option('--exec_list')
@click.option('--savep', default=None)
def dump_clean_lst(exec_list, savep):
final_list = []
for exec_list in exec_list.split(','):
save_dir = osp.dirname(exec_list)
exec_src = exec_list
exec_src_name = osp.basename(exec_src)
srcd = osp.dirname(exec_src)
execsrc2tgt = osp.join(srcd, exec_src_name.split('chunk')[0].rstrip('_')) + '.txt.json'
assert osp.exists(execsrc2tgt)
execsrc2tgt = json2dict(execsrc2tgt)
exec_list = load_exec_list(exec_list)
for srcp in tqdm(exec_list):
tgt_list = execsrc2tgt[srcp]
if srcp not in tgt_list:
tgt_list.append(srcp)
final_list += tgt_list
if savep is None:
savep = osp.join(save_dir, 'cleaned_list.txt')
with open(savep, 'w', encoding='utf8') as f:
f.write('\n'.join(final_list))
@cli.command('dump_complete_lst')
@click.option('--exec_list')
@click.option('--savep', default=None)
def dump_complete_lst(exec_list, savep):
final_list = []
from utils.io_utils import load_exec_list
execp = osp.splitext(exec_list)[0]
savep = execp + '_complete.txt'
for p in load_exec_list(exec_list):
filep = osp.splitext(p)[0] + '_ann.json'
ann = json2dict(filep)
if not ann['cleaned']:
continue
if ann['is_incomplete']:
continue
final_list.append(p)
print(f'filtered: {len(final_list)}')
with open(savep, 'w', encoding='utf8') as f:
f.write('\n'.join(final_list))
@cli.command('update_bodyparsing')
@click.option('--exec_list')
@click.option('--tgt_dir')
@click.option('--parsing_name')
@click.option('--tgt_parsing_name', default='body_parsing.json')
def update_bodyparsing(exec_list, tgt_dir, parsing_name, tgt_parsing_name):
def _check_tag_valid(tag):
valid = True
if 'tag_valid' in metadata_tgt:
if tag in metadata_tgt['tag_valid']:
valid = metadata_tgt['tag_valid'][tag]
else:
valid = False
if not valid:
for k, v in src['drawable'].items():
if v == tag:
valid = True
break
return valid
if tgt_parsing_name is None:
tgt_parsing_name = parsing_name
exec_src = exec_list
exec_src_name = osp.basename(exec_src)
srcd = osp.dirname(exec_src)
execsrc2tgt = osp.join(srcd, exec_src_name.split('chunk')[0].rstrip('_')) + '.txt.json'
assert osp.exists(execsrc2tgt), f'{execsrc2tgt} does not exist!'
execsrc2tgt = json2dict(execsrc2tgt)
exec_list = load_exec_list(exec_list)
n_updates = 0
not_exists = 0
for srcp in tqdm(exec_list):
tgt_list = execsrc2tgt[srcp]
if srcp not in tgt_list:
tgt_list.append(srcp)
relsrcp = osp.relpath(srcp, tgt_dir)
srcp = osp.join(srcd, relsrcp)
src_parsing = osp.join(srcp, parsing_name)
if not osp.exists(srcp):
print(srcp)
not_exists += 1
continue
src = json2dict(src_parsing)
metadata_src = src['metadata']
# metadata_src['is_valid'] = True
if 'tag_valid' not in metadata_src:
metadata_src['tag_valid'] = {}
metadata_src['tag_valid']['objects'] = True
metadata_src['tag_valid']['footwear'] = True
for tgtp in tgt_list:
tgt_parsing = osp.join(tgtp, tgt_parsing_name)
# if osp.exists(tgt_parsing):
# tgt = json2dict(tgt_parsing)
# footwear_valid = True
# metadata_tgt = tgt.get('metadata', {})
# footwear_valid = _check_tag_valid('footwear')
# metadata_src['tag_valid']['footwear'] = footwear_valid
dict2json(src, tgt_parsing)
n_updates += 1
print(not_exists)
@cli.command('render_body_samples')
@click.option('--exec_list')
@click.option('--bg_list')
@click.option('--mask_name', default='bodyparsingv3.json')
@click.option('--save_dir', default='')
@click.option('--rank_to_worldsize', default='', type=str)
@click.option('--save_suffix', default='.png', type=str)
def render_body_samples(exec_list, bg_list, mask_name, save_dir, rank_to_worldsize, save_suffix):
from live2d.scrap_model import animal_ear_detected, Drawable, ImageProcessor, compose_from_drawables, VALID_BODY_PARTS_V3
from utils.cv import fgbg_hist_matching, quantize_image, random_crop, rle2mask, mask2rle, img_alpha_blending, resize_short_side_to, batch_save_masks, batch_load_masks
from utils.torch_utils import seed_everything
seed_everything(42)
hist_match_prob = 0.35
# quantize_prob = 0.25
color_correction_sampler = NameSampler({'hist_match': hist_match_prob, 'quantize': 0.})
exec_list = load_exec_list(exec_list, rank_to_worldsize=rank_to_worldsize)
bg_list = load_exec_list(bg_list)
tagcluster_bodypart = json2dict('assets/tagcluster_bodypart_v2.json')
tag2generaltag = {}
for general_tag, tlist in tagcluster_bodypart.items():
for t in tlist:
if t in tag2generaltag and tag2generaltag[t] != general_tag:
print(f'conflict tag def: {t} - {general_tag}, ' + tag2generaltag[t])
tag2generaltag[t] = general_tag
if save_dir != '':
os.makedirs(save_dir, exist_ok=True)
render_sample = save_dir != ''
MAX_TGT_SIZE = 1280
target_tag_list = VALID_BODY_PARTS_V3 + ['head']
invalid_lst: list[int] = [2094, 1389, 627, 477, 280, 480]
for ii, p in enumerate(tqdm(exec_list)):
try:
lmodel = Live2DScrapModel(p)
load_success = lmodel.load_body_parsing(mask_name)
if not load_success:
print(f'failed to load body parsing, skip: {p}')
continue
metadata = lmodel._body_parsing['metadata']
is_valid = metadata.get('is_valid', True)
is_incomplete = metadata.get('is_incomplete', False)
is_cleaned = metadata.get('cleaned', False)
tag_valid = metadata.get('tag_valid', {})
object_valid = True
foot_valid = True
if not is_valid:
continue
# if is_incomplete:
# continue
# keep_bg = random.random() < 0.3
keep_bg = False
if not is_valid:
continue
valid_drawables: list[Drawable] = []
body_drawables: list[Drawable] = []
h, w = lmodel.size()
x_min, x_max, y_min, y_max = w, 0, h, 0
for d in lmodel.drawables:
d.get_img()
if d.area < 1:
continue
if not keep_bg and d.body_part_tag not in target_tag_list:
continue
valid_drawables.append(d)
if d.body_part_tag in target_tag_list:
body_drawables.append(d)
dxyxy = d.xyxy
x_min = min(x_min, dxyxy[0])
x_max = max(x_max, dxyxy[2])
y_min = min(y_min, dxyxy[1])
y_max = max(y_max, dxyxy[3])
if keep_bg:
x_min = y_min = 0
x_max = w
y_max = h
ch, cw = y_max - y_min, x_max - x_min
scale = min(MAX_TGT_SIZE / max(ch, cw), 1)
nh, nw = ch, cw
if scale < 1:
nh = int(round(nh * scale))
nw = int(round(nw * scale))
new_processor = ImageProcessor(target_frame_size=[nw, nh], crop_bbox=[x_min, y_min, x_max, y_max], pad_to_square=False)
lmodel.final = new_processor(lmodel.final, update_coords_modifiers=True)
lmodel.final_bbox = [
new_processor.crop_bbox[0] + x_min,
new_processor.crop_bbox[1] + y_min,
new_processor.crop_bbox[0] + x_max,
new_processor.crop_bbox[1] + y_max
]
for d in valid_drawables:
d.set_img_processor(new_processor)
d._final_size = [nh, nw]
d.load_img(force_reload=True, img=d.img)
h, w = lmodel.size()
depth_buffer = np.zeros((h, w), dtype=np.uint16)
base_depth = 1
init_drawable_visible_map(valid_drawables)
# part_mask_list, body_final = _compose_body_samples(lmodel)
part_mask_list = []
if not keep_bg:
body_final = lmodel.compose_bodypart_drawables(target_tag_list)
else:
body_final = compose_from_drawables(valid_drawables)
for tag in target_tag_list:
m = lmodel.compose_bodypart_drawables(tag, mask_only=True, final_visible_mask=True).astype(np.uint8)
part_mask_list.append(m)
mask = np.zeros((h, w), dtype=bool)
for d in body_drawables:
m = d.get_full_mask()
if np.any(d.bitwise_and(mask, [0, 0, w, h])):
base_depth += 1
mask = m
else:
mask |= m
d.depth = base_depth
depth_buffer[np.where(m)] = base_depth
depth_dtype = np.uint8
if base_depth > 255:
depth_dtype = np.uint16
depth_buffer = depth_buffer.astype(depth_dtype)
d = osp.abspath(lmodel.directory).replace('\\', '/').rstrip('/').replace('.', '_DOT_')
d1 = d.split('/')[-1]
d2 = d.split('/')[-3]
savename = d2 + '____' + d1
savep = osp.join(save_dir, savename)
masks = part_mask_list
foot_msk_idx = target_tag_list.index('footwear')
object_msk_idx = target_tag_list.index('objects')
leg_msk_idx = target_tag_list.index('legwear')
masks[leg_msk_idx] = masks[leg_msk_idx] | masks[foot_msk_idx]
px = py = 0
final_img = body_final
bgp = random.choice(bg_list)
fh, fw = final_img.shape[:2]
bg = imread(bgp)
fsize = min(max(h, w), MAX_TGT_SIZE)
fsze_max = int(round(fsize * 1.5))
target_bg_size = random.randint(fsize, fsze_max)
bg = resize_short_side_to(bg, target_bg_size)
target_bg_w = target_bg_h = target_bg_size
if fh > fw:
target_bg_w = random.randint(fw, target_bg_size)
elif fw > fh:
target_bg_h = random.randint(fh, target_bg_size)
bg = random_crop(imread(bgp), (target_bg_h, target_bg_w))
px = py = 0
if fh != target_bg_h or fw != target_bg_w:
if fh != target_bg_h:
py = random.randint(0, target_bg_h - fh)
if fw != target_bg_w:
px = random.randint(0, target_bg_w - fw)
blank_final = np.zeros((target_bg_h, target_bg_w, 4), np.uint8)
blank_final[py: py + fh, px: px + fw] = final_img
final_img = blank_final
depth_blank = np.zeros((target_bg_h, target_bg_w), dtype=depth_dtype)
depth_blank[py: py + fh, px: px + fw] = depth_buffer
depth_buffer = depth_blank
for mi, m in enumerate(masks):
blank = np.zeros((target_bg_h, target_bg_w), bool)
blank[py: py + fh, px: px + fw] = m
masks[mi] = blank
fh, fw = final_img.shape[:2]
color_correct = color_correction_sampler.sample()
if color_correct == 'hist_match':
fgbg_hist_matching([final_img], bg, fg_only=True)
wbg = img_alpha_blending([bg, final_img])
wbg[..., -1] = final_img[..., -1]
fh, fw = wbg.shape[:2]
# save_tmp_img(visualize_segs_with_labels(masks, wbg[..., :3], tag_list=target_tag_list, image_weight=0.1))
imwrite(savep, wbg, quality=100, ext=save_suffix)
imwrite(savep + '_depth', depth_buffer, quality=100, ext='.png')
mask_meta_list = [{} for _ in range(len(target_tag_list))] # dont use [{}] * len
mask_meta_list[foot_msk_idx]['is_valid'] = foot_valid
mask_meta_list[object_msk_idx]['is_valid'] = object_valid
batch_save_masks(masks, savep + '.json', mask_meta_list=mask_meta_list)
del masks
# del wbg
del depth_buffer
sample_ann = {'cleaned': is_cleaned, 'is_incomplete': is_incomplete, 'tag_info': {k: {'valid': True, 'exists': False} for k in VALID_BODY_PARTS_V2}, 'final_size': wbg.shape[:2]}
tag_info = sample_ann['tag_info']
# tag_info['footwear']['valid'] = foot_valid
# tag_info['objects']['valid'] = object_valid
for ii, tag in enumerate(target_tag_list):
# if tag == 'footwear' and not foot_valid:
# continue
# if tag == 'objects' and not object_valid:
# continue
if tag == 'head':
drawables = lmodel.get_body_part_drawables(['face', 'irides', 'eyebrow', 'eyewhite', 'eyelash', 'eyewear', 'ears', 'nose', 'mouth'])
else:
drawables = lmodel.get_body_part_drawables(tag)
# if tag == 'legwear':
# drawables += lmodel.get_body_part_drawables('footwear')
drawables = [d for d in drawables if d.area >= 1]
if len(drawables) == 0:
continue
init_drawable_visible_map(drawables)
x_min, x_max, y_min, y_max = fw, 0, fh, 0
for d in drawables:
dxyxy = d.xyxy
x_min = min(x_min, dxyxy[0])
x_max = max(x_max, dxyxy[2])
y_min = min(y_min, dxyxy[1])
y_max = max(y_max, dxyxy[3])
xyxy = [x_min, y_min, x_max, y_max]
dh, dw = y_max - y_min, x_max - x_min
part_final = compose_from_drawables(drawables, xyxy=xyxy)
imwrite(savep + f'_{tag}', part_final, quality=100, ext='.png')
depth_buffer = np.zeros((dh, dw), dtype=depth_dtype)
for d in drawables:
dxyxy = d.xyxy
m = d.final_visible_mask
depth_buffer[dxyxy[1] - y_min: dxyxy[3] - y_min, dxyxy[0] - x_min: dxyxy[2] - x_min][np.where(m)] = d.depth
xyxy = [x_min + px, y_min + py, x_max + px, y_max + py]
imwrite(savep + f'_{tag}_depth', depth_buffer, quality=100, ext='.png')
if tag not in tag_info:
tag_info[tag] = {}
tag_info[tag]['exists'] = True
tag_info[tag]['xyxy'] = xyxy
blank = np.zeros_like(wbg)
blank[xyxy[1]: xyxy[3], xyxy[0]: xyxy[2]] = part_final
# save_tmp_img(wbg)
# save_tmp_img(img_alpha_blending([wbg, blank]))
# pass
dict2json(sample_ann, savep + '_ann.json')
except Exception as e:
# raise
print(f'Failed to process {p}: {e}')
@cli.command('render_simple_samples')
@click.option('--exec_list')
@click.option('--bg_list')
@click.option('--save_dir', default='')
@click.option('--rank_to_worldsize', default='', type=str)
@click.option('--save_suffix', default='.png', type=str)
def render_simple_samples(exec_list, bg_list, save_dir, rank_to_worldsize, save_suffix):
from live2d.scrap_model import animal_ear_detected, Drawable, ImageProcessor, compose_from_drawables
from utils.cv import fgbg_hist_matching, quantize_image, random_crop, rle2mask, mask2rle, img_alpha_blending, resize_short_side_to, batch_save_masks, batch_load_masks
from utils.torch_utils import seed_everything
seed_everything(42)
hist_match_prob = 0.35
# quantize_prob = 0.25
color_correction_sampler = NameSampler({'hist_match': hist_match_prob, 'quantize': 0.})
is_pkl = False
if exec_list.endswith('.pkl'):
src_dir = osp.dirname(exec_list)
import pickle
is_pkl = True
with open(exec_list, 'rb') as f:
exec_list = pickle.load(f)
exec_list = load_exec_list(exec_list, rank_to_worldsize=rank_to_worldsize)
bg_list = load_exec_list(bg_list)
# tagcluster_bodypart = json2dict('assets/tagcluster_bodypart_v2.json')
# tag2generaltag = {}
# for general_tag, tlist in tagcluster_bodypart.items():
# for t in tlist:
# if t in tag2generaltag and tag2generaltag[t] != general_tag:
# print(f'conflict tag def: {t} - {general_tag}, ' + tag2generaltag[t])
# tag2generaltag[t] = general_tag
if save_dir != '':
os.makedirs(save_dir, exist_ok=True)
render_sample = save_dir != ''
MAX_TGT_SIZE = 1024
# valid_taglst = set(list(tag2generaltag.keys()) + ['smile'])
nsaved = 0
# nsaved += len(exec_list)
for ii, p in enumerate(tqdm(exec_list)):
try:
if is_pkl:
tags = p['tags']
p = osp.join(src_dir, p['file_path'])
img = Image.open(p).convert('RGBA')
if not is_pkl:
from annotators.wdv3_tagger import apply_wdv3_tagger
img_input = pil_ensure_rgb(img)
img_input = img_input.resize((448, 448), resample=Image.Resampling.LANCZOS)
caption, taglist, ratings, character, general = apply_wdv3_tagger(img_input)
tags = general
# tags = [tag for tag in tags if tag in valid_taglst]
img = np.array(img)
x1, y1, x2, y2 = cv2.boundingRect(cv2.findNonZero((img[..., -1] > 25).astype(np.uint8)))
x2 += x1
y2 += y1
img = img[y1: y2, x1: x2].copy()
ch, cw = img.shape[:2]
scale = min(MAX_TGT_SIZE / max(ch, cw), 1)
nh, nw = ch, cw
if scale < 1:
nh = int(round(nh * scale))
nw = int(round(nw * scale))
img = cv2.resize(img, dsize=(nw, nh), interpolation=cv2.INTER_AREA)
savename = str(nsaved).zfill(5)
savep = osp.join(save_dir, savename)
h, w = img.shape[:2]
final_img = img
bgp = random.choice(bg_list)
fh, fw = final_img.shape[:2]
bg = imread(bgp)
fsize = min(max(h, w), MAX_TGT_SIZE)
fsze_max = int(round(fsize * 1.5))
target_bg_size = random.randint(fsize, fsze_max)
bg = resize_short_side_to(bg, target_bg_size)
target_bg_w = target_bg_h = target_bg_size
if fh > fw:
target_bg_w = random.randint(fw, target_bg_size)
elif fw > fh:
target_bg_h = random.randint(fh, target_bg_size)
bg = random_crop(imread(bgp), (target_bg_h, target_bg_w))
px = py = 0
if fh != target_bg_h or fw != target_bg_w:
if fh != target_bg_h:
py = random.randint(0, target_bg_h - fh)
if fw != target_bg_w:
px = random.randint(0, target_bg_w - fw)
blank_final = np.zeros((target_bg_h, target_bg_w, 4), np.uint8)
blank_final[py: py + fh, px: px + fw] = final_img
final_img = blank_final
fh, fw = final_img.shape[:2]
color_correct = color_correction_sampler.sample()
if color_correct == 'hist_match':
fgbg_hist_matching([final_img], bg, fg_only=True)
wbg = img_alpha_blending([bg, final_img])
wbg[..., -1] = final_img[..., -1]
fh, fw = wbg.shape[:2]
# save_tmp_img(visualize_segs_with_labels(masks, wbg[..., :3], tag_list=target_tag_list, image_weight=0.1))
imwrite(savep, wbg, quality=100, ext=save_suffix)
with open(savep + '.txt', 'w', encoding='utf8') as f:
f.write(','.join(tags))
nsaved += 1
except Exception as e:
# raise
print(f'Failed to process {p}: {e}')
if __name__ == '__main__':
cli()