-
Notifications
You must be signed in to change notification settings - Fork 3
/
IVOS_main_youtube.py
399 lines (319 loc) · 19.3 KB
/
IVOS_main_youtube.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
from davisinteractive.dataset import Davis
from libs import custom_transforms as tr
import os
import numpy as np
import json
from PIL import Image
import csv
from datetime import datetime
import cv2
import torch
from torch.autograd import Variable
from torchvision import transforms
from torch.utils.data import DataLoader
import torch.nn.functional as F
from libs import utils_custom, utils_visualize
from config import Config
from networks.network import NET_GAmap
from libs.youtubesession import YoutubeSession
from datasets_torch.torchdatasets_youtube_test import YoutubeVOS
class Main_tester(object):
def __init__(self, config, n_total_rounds=4):
self.config = config
self.Davisclass = Davis(self.config.davis_dataset_dir)
self.current_time = datetime.now().strftime('%Y%m%d-%H%M%S')
self._palette = Image.open(self.config.palette_dir).getpalette()
self.save_res_dir = str()
self.save_log_dir = str()
self.save_logger = None
self.save_csvsummary_dir = str()
self.net = NET_GAmap()
self.net.cuda()
self.net.eval()
self.net.load_state_dict(torch.load('checkpoints/ckpt_wo_YTIVOStrainset.pth'))
self.max_nb_interactions = 4
self.img_size, self.num_frames, self.n_objects, self.final_masks, self.tmpdict_siact = None, None, None, None, None
self.pad_info, self.hpad1, self.wpad1, self.hpad2, self.wpad2 = None, None, None, None, None
self.n_total_rounds = n_total_rounds
def run_youtube(self, n_ipoint, save_res_dir):
self.n_init_points = n_ipoint
test_load_points_dir = 'etc/0youtube_iPoints/point_info_{:02d}.json'.format(n_ipoint)
self.session = YoutubeSession(self.config, n_total_rounds=self.n_total_rounds, test_load_points_dir= test_load_points_dir)
self.save_res_dir = save_res_dir + '/iPoint{:02d}'.format(n_ipoint)
utils_custom.mkdir(self.save_res_dir)
self.save_csvsummary_dir = os.path.join(self.save_res_dir, 'summary_in_csv.csv')
self.save_log_dir = os.path.join(self.save_res_dir, 'test_logs.txt')
self.save_logger = utils_custom.logger(self.save_log_dir)
self.save_logger.printNlog(dir_name)
with torch.no_grad():
pf = self.run_IVOS()
return pf
def run_IVOS(self):
seen_seq = {}
output_dict = dict()
output_dict['average_objs_iou'] = dict()
output_dict['average_iact_iou'] = np.zeros(self.max_nb_interactions)
output_dict['annotated_frames'] = dict()
with open(self.save_csvsummary_dir, mode='a') as csv_file:
writer = csv.writer(csv_file, delimiter=',', quotechar='"', quoting=csv.QUOTE_MINIMAL)
writer.writerow(['sequence', 'obj_idx'] + ['round-' + str(i + 1) for i in range(self.max_nb_interactions)])
for ii,video in enumerate(self.session.youtube_val_videos):
print('{:04d}th-video'.format(ii+1))
self.img_size, self.num_frames, self.n_objects = \
self.session.initialize_session_for_video(video)
self.sequence = video
anno_dict = {'frames': [], 'annotated_masks': [], 'masks_tobe_modified': []}
seen_seq[self.sequence] = 1 if self.sequence not in seen_seq.keys() else seen_seq[self.sequence] + 1
scr_id = seen_seq[self.sequence]
self.final_masks = np.zeros([self.num_frames, self.img_size[0], self.img_size[1]])
pd_r_img, self.pad_info = utils_custom.apply_pad(self.resize_shorter480_seg(self.final_masks[0]))
self.hpad1, self.wpad1 = self.pad_info[0][0], self.pad_info[1][0]
self.hpad2, self.wpad2 = self.pad_info[0][1], self.pad_info[1][1]
self.prob_map_of_frames = torch.zeros((self.num_frames, self.n_objects + 1, pd_r_img.shape[0], pd_r_img.shape[1])).cuda()
self.anno_6chEnc_r4_list = []
self.anno_3chEnc_r4_list = []
self.scores_ni_nf = np.zeros([8, self.num_frames])
IoU_over_eobj = []
for n_interaction in range(1, self.n_total_rounds+1):
annotated_now = self.session.prev_frames_used_in_index[-1]
anno_dict['frames'].append(annotated_now) # Where we save annotated frames
anno_dict['masks_tobe_modified'].append(self.final_masks[annotated_now]) # mask before modefied at the annotated frame
if n_interaction == 1:
data_points_or_scr = self.session.get_points_firstround()
else:
data_points_or_scr = self.session.get_scrdata_currentround()
self.save_logger.printNlog('\nRunning sequence {} in round: {}'.format(self.sequence, n_interaction))
# Get Predicted mask & Mask decision from pred_mask
self.final_masks = self.run_VOS_singleiact(n_interaction, data_points_or_scr, anno_dict['frames']) # self.final_mask changes
# Limit candidate frames
if n_interaction != self.max_nb_interactions:
self.scores_ni_nf[n_interaction] = self.scores_ni_nf[n_interaction-1]
current_score_np = self.scores_ni_nf[n_interaction-1]
if self.config.test_guide_method=='RS1':
next_scribble_frame_candidates = list(np.argsort(current_score_np)[:1])
elif self.config.test_guide_method=='RS4':
sorted_score_idx = np.argsort(current_score_np)
exclude_range = self.num_frames/10
excluded_next_candidates = []
next_scribble_frame_candidates = []
for i in range(self.num_frames):
if not sorted_score_idx[i] in excluded_next_candidates:
next_scribble_frame_candidates.append(sorted_score_idx[i])
excluded_next_candidates += list(range(
int(sorted_score_idx[i]-(exclude_range/2)+0.5), int(sorted_score_idx[i]+(exclude_range/2)+0.5)))
if len(next_scribble_frame_candidates)==4:
break
elif self.config.test_guide_method=='wo_RS':
next_scribble_frame_candidates=None
else:
raise NotImplementedError
# Submit your prediction
self.session.submit_masks(self.final_masks, next_scribble_frame_candidates) # F, H, W
if self.config.test_save_pngs_option:
utils_custom.mkdir(os.path.join(self.save_res_dir, 'result_video', '{}/round{:02d}'.format(self.sequence, n_interaction)))
for fr_idx in range(self.num_frames):
savefname = os.path.join(self.save_res_dir, 'result_video','{}/round{:02d}'.format(self.sequence, n_interaction),self.session.frame_list[fr_idx])
tmpPIL = Image.fromarray(self.final_masks[fr_idx].astype(np.uint8), 'P')
tmpPIL.putpalette(self._palette)
tmpPIL.save(savefname)
## Visualizers and Saver
# IoU estimation
IoU_over_eobj.append(self.session.metric)
# save final mask in anno_dict
anno_dict['annotated_masks'].append(self.final_masks[annotated_now]) # mask after modefied at the annotated frame
if n_interaction == 4: # After Lastround -> total 90 iter
# IoU manager
IoU_over_eobj = np.stack(IoU_over_eobj, axis=0) # niact,frames,n_obj
IoUeveryround_perobj = np.mean(IoU_over_eobj, axis=1) # niact,n_obj
output_dict['average_iact_iou'] += np.sum(IoU_over_eobj[list(range(n_interaction)), anno_dict['frames']], axis=-1)
output_dict['annotated_frames'][self.sequence] = anno_dict['frames']
if self.config.test_save_pngs_option:
savefiledir = os.path.join(self.save_res_dir, 'plot_IoU_perObj')
utils_custom.mkdir(savefiledir)
for obj_idx in range(self.n_objects):
savefilename = os.path.join(savefiledir, self.sequence + '-obj' + str(obj_idx + 1) + '_first{:03d}final{:03d}.png'
.format(int(1000 * IoUeveryround_perobj[0, obj_idx]),
int(1000 * IoUeveryround_perobj[-1, obj_idx])))
utils_visualize.visualize_interactionIoU(IoU_over_eobj[:, :, obj_idx], self.sequence + '-obj' + str(obj_idx + 1),
anno_dict['frames'], save_dir=savefilename, show_propagated_region=True)
# write csv
for obj_idx in range(self.n_objects):
with open(self.save_csvsummary_dir, mode='a') as csv_file:
writer = csv.writer(csv_file, delimiter=',', quotechar='"', quoting=csv.QUOTE_MINIMAL)
writer.writerow([self.sequence, str(obj_idx + 1)] + list(IoUeveryround_perobj[:, obj_idx]))
savefname = self.save_res_dir + '/summary.json'
performanceJF = list(self.session.performance_summary())
with open(savefname, "w") as json_file:
json.dump(performanceJF, json_file)
torch.cuda.empty_cache()
model = None
self.save_logger.printNlog(str(performanceJF))
return performanceJF
def run_VOS_singleiact(self, n_interaction, data_points_or_scr, annotated_frames):
'''
pm_ps_ns => # n_obj,3,h,w
'''
annotated_frames_np = np.array(annotated_frames)
annotated_now = annotated_frames[-1]
prop_list = utils_custom.get_prop_list(annotated_frames, annotated_now, self.num_frames, proportion=self.config.test_propagation_proportion)
if n_interaction == 1:
pm_ps_ns_3ch_np = self.point_data_to_img(data_points_or_scr)
else:
pm_ps_ns_3ch_np = self.scr_data_to_img(data_points_or_scr, n_interaction, annotated_now)
pm_ps_ns_3ch_t = torch.from_numpy(pm_ps_ns_3ch_np).cuda()
if (prop_list[0] != annotated_now) and (prop_list.count(annotated_now) != 2):
print(str(prop_list))
raise NotImplementedError
print(str(prop_list)) # we made our proplist first backward, and then forward
composed_transforms = transforms.Compose([tr.Normalize_ApplymeanvarImage(self.config.mean, self.config.var),
tr.ToTensor()])
db_test = YoutubeVOS(self.config.youtube_dataset_dir, self.config, transform=composed_transforms, custom_frames=prop_list,
seq_name=self.sequence, resize=True,)
testloader = DataLoader(db_test, batch_size=1, shuffle=False, num_workers=self.config.test_n_workers, pin_memory=True)
flag = 0 # 1: propagating backward, 2: propagating forward
print('[{:01d} round] processing...'.format(n_interaction))
for ii, batched in enumerate(testloader):
# batched : image, scr_img, 0~fr, meta
inpdict = dict()
operating_frame = int(batched['meta']['frame_id'][0])
for inp in batched:
if inp == 'meta': continue
inpdict[inp] = Variable(batched[inp]).cuda()
inpdict['image'] = inpdict['image'].expand(self.n_objects, -1, -1, -1)
#################### Iaction ########################
if operating_frame == annotated_now: # Check the round is on interaction
if flag == 0:
flag += 1
adjacent_to_anno = True
elif flag == 1:
flag += 1
adjacent_to_anno = True
continue
else:
raise NotImplementedError
pm_ps_ns_3ch_t = torch.nn.ReflectionPad2d(self.pad_info[1] + self.pad_info[0])(pm_ps_ns_3ch_t)
inputs = torch.cat([inpdict['image'], pm_ps_ns_3ch_t], dim=1)
anno_3chEnc_r4, _ = self.net.encoder_3ch.forward(inpdict['image'])
neighbor_pred_onehot_sal, anno_6chEnc_r4 = self.net.forward_obj_feature_extractor(inputs) # [nobj, 1, P_H, P_W], # [n_obj,2048,h/16,w/16]
output_logit, r4_anno, score = self.net.forward_prop(
[anno_3chEnc_r4], inpdict['image'], [anno_6chEnc_r4],
anno_3chEnc_r4, torch.sigmoid(neighbor_pred_onehot_sal),
anno_fr_list= annotated_frames_np, que_fr= operating_frame) # [nobj, 1, P_H, P_W]
output_prob_tmp = F.softmax(output_logit, dim=1) # [nobj, 2, P_H, P_W]
output_prob_tmp = output_prob_tmp[:, 1] # [nobj, P_H, P_W]
one_hot_outputs_t = F.softmax(self.soft_aggregation(output_prob_tmp), dim=0) # [nobj+1, P_H, P_W]
anno_onehot_prob = one_hot_outputs_t.clone()[1:].unsqueeze(1) # [nobj, 1, P_H, P_W]
anno_3chEnc_r4, r2_prev_fromanno = self.net.encoder_3ch.forward(inpdict['image'])
self.anno_6chEnc_r4_list.append(anno_6chEnc_r4)
self.anno_3chEnc_r4_list.append(anno_3chEnc_r4)
if len(self.anno_6chEnc_r4_list) != len(annotated_frames):
raise NotImplementedError
#################### Propagation ########################
else:
# Flag [1: propagating backward, 2: propagating forward]
if adjacent_to_anno:
r4_neighbor = r4_anno
neighbor_pred_onehot = anno_onehot_prob
else:
r4_neighbor = r4_que
neighbor_pred_onehot = targ_onehot_prob
adjacent_to_anno = False
output_logit, r4_que, score = self.net.forward_prop(
self.anno_3chEnc_r4_list, inpdict['image'], self.anno_6chEnc_r4_list,
r4_neighbor, neighbor_pred_onehot,
anno_fr_list= annotated_frames_np, que_fr= operating_frame) # [nobj, 1, P_H, P_W]
output_prob_tmp = F.softmax(output_logit, dim=1) # [nobj, 2, P_H, P_W]
output_prob_tmp = output_prob_tmp[:, 1] # [nobj, P_H, P_W]
one_hot_outputs_t = F.softmax(self.soft_aggregation(output_prob_tmp), dim=0) # [nobj+1, P_H, P_W]
targ_onehot_prob = one_hot_outputs_t.clone()[1:].unsqueeze(1) # [nobj, 1, P_H, P_W]
# Final mask indexing
self.prob_map_of_frames[operating_frame] = one_hot_outputs_t
onehot_out_tmp = F.interpolate(
one_hot_outputs_t[:,self.hpad1:-self.hpad2, self.wpad1:-self.wpad2].unsqueeze(dim=0), size=self.final_masks[0].shape)
self.final_masks[operating_frame] = torch.argmax(onehot_out_tmp[0],dim=0).cpu().numpy().astype(np.uint8)
self.scores_ni_nf[n_interaction-1,operating_frame] = score
torch.cuda.empty_cache()
return self.final_masks
def soft_aggregation(self, ps):
num_objects, H, W = ps.shape
em = torch.zeros(num_objects +1, H, W).cuda()
em[0] = torch.prod(1-ps, dim=0) # bg prob
em[1:num_objects+1] = ps # obj prob
em = torch.clamp(em, 1e-7, 1-1e-7)
logit = torch.log((em /(1-em)))
return logit
def resize_shorter480(self, img,seg):
ori_h, ori_w = img.shape[0], img.shape[1]
if ori_w >= ori_h:
if ori_h ==480:
return img,seg
new_h = 480
new_w = int((ori_w / ori_h) * 480)
else:
if ori_w ==480:
return img,seg
new_w = 480
new_h = int((ori_h / ori_w) * 480)
output_size = (new_w, new_h)
new_img = cv2.resize(img, output_size, interpolation=cv2.INTER_CUBIC)
new_seg = cv2.resize(seg, output_size, interpolation=cv2.INTER_NEAREST)
return new_img, new_seg
def resize_shorter480_seg(self, seg):
ori_h, ori_w = seg.shape[0], seg.shape[1]
if ori_w >= ori_h:
new_h = 480
new_w = int((ori_w / ori_h) * 480)
else:
new_w = 480
new_h = int((ori_h / ori_w) * 480)
output_size = (new_w, new_h)
new_seg = cv2.resize(seg, output_size, interpolation=cv2.INTER_NEAREST)
return new_seg
def point_data_to_img(self, points_data):
zeros_map = np.zeros_like(self.resize_shorter480_seg(self.final_masks[0]))
ori_h, ori_w = self.final_masks[0].shape
if ori_w >= ori_h: sizerate = 480 / ori_h
else: sizerate = 480 / ori_w
try:points_data = (np.asarray(points_data)*sizerate).astype(np.int64)
except:
a=1
pm_ps_ns_3ch_t = [] # n_obj,3,h,w
for obj_id in range(1, self.n_objects + 1):
ptmap_tmp = np.zeros_like(zeros_map, dtype=np.float32)
ptmap_tmp[points_data[obj_id-1][0, :], points_data[obj_id-1][1, :]]=1
pos_ptimg = utils_custom.scrimg_postprocess(ptmap_tmp, dilation=7, blur=True, blursize=(5, 5), var=6.0)
pm_ps_ns_3ch_t.append(np.stack([np.ones_like(pos_ptimg) / 2, pos_ptimg, np.zeros_like(pos_ptimg)], axis=0))
pm_ps_ns_3ch_t = np.stack(pm_ps_ns_3ch_t, axis=0) # n_obj,3,h,w
return pm_ps_ns_3ch_t
def scr_data_to_img(self, scribbles_data, n_interaction, annotated_now):
prev_mask = self.resize_shorter480_seg(self.final_masks[annotated_now])
scr_data = scribbles_data['scribbles']
# Interaction settings
pm_ps_ns_3ch_t = [] # n_obj,3,h,w
for obj_id in range(1, self.n_objects + 1):
prev_round_input = (prev_mask == obj_id).astype(np.float32) # H,W
pos_scrimg, neg_scrimg = utils_custom.scribble_to_image(scr_data, annotated_now, obj_id,
dilation=self.config.scribble_dilation_param,
prev_mask=prev_mask, blur=True,
singleimg=False, seperate_pos_neg=True)
pm_ps_ns_3ch_t.append(np.stack([prev_round_input, pos_scrimg, neg_scrimg], axis=0))
pm_ps_ns_3ch_t = np.stack(pm_ps_ns_3ch_t, axis=0) # n_obj,3,h,w
return pm_ps_ns_3ch_t
if __name__ == '__main__':
config = Config()
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = str(config.test_gpu_id)
current_time = datetime.now().strftime('%Y%m%d-%H%M%S')
dir_name = os.path.split(os.path.split(__file__)[0])[1] + '[JF]_[' + config.test_guide_method + ']_' + current_time
save_res_dir = os.path.join(config.test_result_yt_dir, dir_name)
utils_custom.mkdir(save_res_dir)
tester = Main_tester(config)
performanceJF_05 = tester.run_youtube(5, save_res_dir)
performanceJF_10 = tester.run_youtube(10, save_res_dir)
performanceJF_20 = tester.run_youtube(20, save_res_dir)
performanceJF_50 = tester.run_youtube(50, save_res_dir)
savefname = save_res_dir + '/summary_total.json'
performanceJF = np.array(performanceJF_05)+np.array(performanceJF_10)+np.array(performanceJF_20)+np.array(performanceJF_50)
performanceJF = performanceJF.tolist()
with open(savefname, "w") as json_file:
json.dump(performanceJF, json_file)