-
Notifications
You must be signed in to change notification settings - Fork 0
/
SegMyO_selection.py
282 lines (246 loc) · 14.3 KB
/
SegMyO_selection.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Select the best segment in a list of proposals (the output of a segmentation model),
for a given bounding box and/or a given label, by using a combination of several
geometric criteria on the bounding box (or the image if no bounding box is given),
and a semantic criterion if the expected label is given.
based on segmentation models from:
- https://pytorch.org/docs/stable/torchvision/models.html
author: Robin Deléarde @LIPADE (Université de Paris)
date: 23/12/2020
"""
#%% import libraries
import numpy as np
#from skimage.morphology import area_opening # used to remove small regions in the segmentation outputs (optional)
from sematch.semantic.similarity import WordNetSimilarity # WordNet similarity with sematch (https://github.com/gsi-upm/sematch)
from test_utils import compute_iou_dice
wns = WordNetSimilarity()
#%% labels for pytorch/torchvision models trained on COCO or PASCAL VOC (https://pytorch.org/docs/stable/torchvision/models.html)
# labels for torchvision Mask R-CNN and Faster R-CNN {detection + instance segmentation} models, pre-trained on COCO (80 classes)
COCO_labels = ['__background__', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
'train', 'truck', 'boat', 'traffic light', 'fire hydrant', 'N/A', 'stop sign',
'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow',
'elephant', 'bear', 'zebra', 'giraffe', 'N/A', 'backpack', 'umbrella', 'N/A', 'N/A',
'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball',
'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket',
'bottle', 'N/A', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl',
'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza',
'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', 'N/A', 'dining table',
'N/A', 'N/A', 'toilet', 'N/A', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone',
'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'N/A', 'book',
'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush']
# labels for torchvision FCN and DeepLabV3 segmentation models, pre-trained on COCO subset with Pascal_VOC classes (20 classes)
Pascal_VOC_labels = ['__background__', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus',
'car', 'cat', 'chair', 'cow', 'dining table', 'dog', 'horse', 'motorbike',
'person', 'potted plant', 'sheep', 'sofa', 'train', 'tv monitor']
#%% compute criteria
def compute_criteria(seg_model, seg_output, proposal_nr, item_bbox=None, item_label=None):
"""
Compute the criteria scores for a given bounding box and/or a given label,
for a given region/segment proposal in a segmentation output.
args:
seg_output: segmentation output dictionary structure with attributes:
'labels': list of predicted labels
'scores': list of segmentation scores
'masks': list of segmentation masks
item_label: label of the expected object (if None, use geometric criteria only)
item_bbox: bbox coordinates of the expected object in the original image, in format
array([x_start, x_stop, y_start, y_stop]), None will take the full image
output:
1D array containing the scores for each criterion, in the following order:
[seg_score, dist_to_edges_score, bbox_area_ratio, item_r_area, dist_to_center_score, semantic_score, my_score1, my_score2]
"""
current_item = seg_output['labels'][proposal_nr]
seg_score = seg_output['scores'][proposal_nr] # criterion 0: segmentation score
mask = seg_output['masks'][proposal_nr]
if seg_model in ['Mask R-CNN', 'Faster R-CNN', 'FCN', 'DeepLabV3']:
seg_score = seg_score.item()
mask = mask[0].detach().numpy()
if seg_model in ['Mask R-CNN', 'Faster R-CNN']:
current_item = COCO_labels[current_item.item()]
elif seg_model in ['FCN', 'DeepLabV3']:
current_item = Pascal_VOC_labels[current_item.item()]
else:
raise Exception('segmentation model not implemented')
# plt.imshow(mask); plt.show()
# mask = area_opening(mask, max(400, mask.shape[0]*mask.shape[1]//135)) # remove small regions (optional)
# plt.imshow(mask); plt.show()
mask_area = mask.sum()
if item_bbox is None:
mask_bbox = mask
item_bbox = np.array([0, mask.shape[0], 0, mask.shape[1]], dtype=np.uint8)
else:
mask_bbox = mask[item_bbox[0]:item_bbox[1], item_bbox[2]:item_bbox[3]]
mask_bbox_area = mask_bbox.sum()
if mask_bbox_area==0:
return np.zeros((8)) # set all scores to 0 if item is oustide bbox
else:
# geometric criterion 1: distance to the edges of the bbox -> basic necessary and universal criterion
x_bbox,y_bbox = np.nonzero(mask_bbox)
if len(x_bbox)>0: # equivalent to mask_bbox_area>0
dist_up = min(x_bbox)/mask_bbox.shape[0]
dist_down = (mask_bbox.shape[0]-1-max(x_bbox))/mask_bbox.shape[0]
dist_left = min(y_bbox)/mask_bbox.shape[1]
dist_right = (mask_bbox.shape[1]-1-max(y_bbox))/mask_bbox.shape[1]
max_distance = max(dist_up, dist_down, dist_left, dist_right)
else:
max_distance = 1.
dist_to_edges_score = 1.-max_distance
# geometric criterion 2: bbox_area_ratio (NB: bad if the object is not a single instance)
bbox_area_ratio = mask_bbox_area/mask_area
# geometric criterion 3: item relative area -> to put more weight to large objects (relatively to the bbox size)
item_r_area = mask_bbox_area/mask_bbox.shape[0]/mask_bbox.shape[1]
# geometric criterion 4: distance to the center of the bounding box -> only for barycentric objects or to avoid objects only on the edges
center_x, center_y = item_bbox[0]+mask_bbox.shape[0]//2, item_bbox[2]+mask_bbox.shape[1]//2 # center of bbox/image
x,y = np.nonzero(mask)
# dist = np.sqrt((x-center_x)**2+(y-center_y)**2)
dist = abs(x-center_x)/mask_bbox.shape[0]+abs(y-center_y)/mask_bbox.shape[1]
dist_norm = 1-np.exp(-dist)
mask_values = mask[x,y]
dist_to_center_score = 1.-(mask_values*dist_norm).sum()/mask_area # NB: mask_bbox_area=mask_values.sum()
# semantic criterion: label matching
if item_label is not None and current_item is not None:
if current_item.replace(' ', '_')==item_label.replace(' ', '_'):
semantic_score = 1.
else:
semantic_score = wns.word_similarity(current_item.replace(' ', '_'), item_label.replace(' ', '_'), 'wpath')
# 3th argument = method, among: 'li', 'lin', 'wup', 'res', 'jcn', 'wpath' (cf. https://gsi-upm.github.io/sematch/similarity/)
# global criterion
if item_label is not None and semantic_score>0:
my_score1 = (seg_score + 2*bbox_area_ratio + 2*dist_to_edges_score + item_r_area + dist_to_center_score + 5*semantic_score)/12
else:
my_score1 = (seg_score + 2*bbox_area_ratio + 2*dist_to_edges_score + item_r_area + dist_to_center_score)/7
if seg_model in ['Mask R-CNN', 'Faster R-CNN']: # instance segmentation => bbox_area_ratio should be close to 1
my_score2 = min(seg_score, bbox_area_ratio, dist_to_edges_score, my_score1)
else:
my_score2 = min(seg_score, dist_to_edges_score, my_score1)
return np.array([seg_score, dist_to_edges_score, bbox_area_ratio, item_r_area, dist_to_center_score, semantic_score, my_score1, my_score2])
def compute_criteria_all_proposals(seg_model, seg_output, item_bbox=None, item_label=None, verbose=True):
"""
Compute the criteria scores for a given bounding box and/or a given label,
for all region/segment proposals in a segmentation output.
args:
seg_output: segmentation outputs dictionary structure with attributes:
'labels': list of predicted labels
'scores': list of segmentation scores
'masks': list of segmentation masks
item_bbox: bbox coordinates of the expected object in the original image, in format
array([x_start, x_stop, y_start, y_stop]), None will take the full image
item_label: label of the expected object (if None, use geometric criteria only)
output:
2D array containing the scores for each proposal (lines) and each criterion (columns), in the following order:
[seg_score, dist_to_edges_score, bbox_area_ratio, item_r_area, dist_to_center_score, semantic_score, my_score1, my_score2])
"""
nb_items = len(seg_output['scores'])
if nb_items==0:
if verbose:
print('Did not find any item in segmentation outputs')
return None
else:
scores = np.zeros((nb_items, 8))
for i in range(nb_items):
scores[i] = compute_criteria(seg_model, seg_output, i, item_bbox, item_label)
return np.transpose(scores)
#%% select the best output
def select_output(img_size, seg_model, seg_output, scores_list, item_bbox=None, item_label=None, score_thresh=0.25, bin_thresh=0.3, remove_outside_bbox=True, verbose=True):
"""
Select the best region/segment proposal in a segmentation output
for a given bounding box and/or a given label,
based on a list of criteria scores computed previously.
args:
seg_output: segmentation outputs dictionary structure with attributes:
'labels': list of predicted labels
'scores': list of segmentation scores
'masks': list of segmentation masks
scores_list: 1D array of scores for all proposals (1 score/proposal)
item_bbox: bbox coordinates of the expected object in the original image, in format
array([x_start, x_stop, y_start, y_stop]), None will take the full image
item_label: label of the expected object (if None, use geometric criteria only)
score_thresh: threshold for the score, return None output if score<score_thresh for all items
bin_thresh: threshold for the binarization of the segmentation mask (no binarization if None)
outputs:
mask: mask of the selected output (binary mask if bin_thresh is not None)
predicted_label: label of the selected output
my_score: score of the selected output
"""
nb_items = len(seg_output['scores'])
if nb_items==0:
if verbose:
print('Did not find any item in segmentation outputs')
empty_img = np.zeros(img_size, dtype=np.uint8)
return empty_img, None, 0.
else:
# # uncomment to compute scores inside selection (NB: scores are not returned here)
# criteria_scores = compute_criteria_all_proposals(seg_model, seg_output, item_bbox, item_label, verbose)
# scores_list = criteria_scores[criterion_nr]
my_score = np.max(scores_list)
i = np.argmax(scores_list)
if score_thresh is not None and my_score<score_thresh:
if verbose:
print('Did not find any item with score > score_tresh')
empty_img = np.zeros(img_size, dtype=np.uint8)
return empty_img, None, 0.
else:
if seg_model in ['Mask R-CNN', 'Faster R-CNN']:
predicted_label = COCO_labels[seg_output['labels'][i].item()]
elif seg_model in ['FCN', 'DeepLabV3']:
predicted_label = Pascal_VOC_labels[seg_output['labels'][i].item()]
if verbose:
print('Found item "'+str(predicted_label)+'" with score %.2f' %my_score)
# extract mask image
mask = seg_output['masks'][i]
if seg_model in ['Mask R-CNN', 'Faster R-CNN', 'FCN', 'DeepLabV3']:
mask = mask[0].detach().numpy()
# plt.imshow(mask); plt.show()
if bin_thresh is not None:
mask = (mask>=bin_thresh).astype(np.uint8) # binarization of the segmentation
# plt.imshow(img_seg); plt.show()
# remove part outside of the bounding box (optional)
if remove_outside_bbox and item_bbox is not None:
mask_bbox = mask[item_bbox[0]:item_bbox[1], item_bbox[2]:item_bbox[3]]
if bin_thresh is not None:
mask = np.zeros(img_size, dtype=np.uint8)
else:
mask = np.zeros(img_size, dtype=np.float32)
mask[item_bbox[0]:item_bbox[1], item_bbox[2]:item_bbox[3]] = mask_bbox
# plt.imshow(mask); plt.show()
return mask, predicted_label, my_score
#%% compute mIoU for each mask with given ground truth
def compute_mIoUs(seg_output, seg_model, mask_gt, item_bbox, seg_thresh=0.25, bin_thresh=0.3, verbose=True):
"""
compute mIoU for each mask in seg_output with seg_score>seg_thresh, with mask_gt as ground truth
only with bbox_before = False
"""
nb_items = len(seg_output['scores'])
if nb_items==0:
if verbose:
print('Did not find any item in segmentation outputs')
return None
else:
seg_score = seg_output['scores'][0]
if seg_model in ['Mask R-CNN', 'Faster R-CNN', 'FCN', 'DeepLabV3']:
seg_score = seg_score.item()
if seg_score<seg_thresh:
if verbose:
print('Segmentation score < threshold for all items')
return None
else:
i=0
mIoU_list = []
while i<nb_items and seg_score>=seg_thresh:
seg_score = seg_output['scores'][i] # criterion 1: segmentation score (in [0,1])
if seg_model in ['Mask R-CNN', 'Faster R-CNN', 'FCN', 'DeepLabV3']:
seg_score = seg_score.item()
if seg_score>=seg_thresh: # necessary for last item in while loop
mask = seg_output['masks'][i]
if seg_model in ['Mask R-CNN', 'Faster R-CNN', 'FCN', 'DeepLabV3']:
mask = mask[0].detach().numpy()
mask = (mask>=bin_thresh).astype(np.uint8) # binarization of the segmentation
img_seg = mask[item_bbox[0]:item_bbox[1], item_bbox[2]:item_bbox[3]]
img_full = np.zeros(mask.shape)
img_full[item_bbox[0]:item_bbox[1], item_bbox[2]:item_bbox[3]] = img_seg
iou_score, dice_score = compute_iou_dice(img_full, mask_gt)
mIoU_list.append(iou_score)
i+=1
return mIoU_list