-
Notifications
You must be signed in to change notification settings - Fork 52
/
backgroundFalsePosErrors.py
222 lines (192 loc) · 9.75 KB
/
backgroundFalsePosErrors.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
## imports
import os, time
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.path as mplPath
from scipy.misc import imresize
import skimage.io as io
# package imports
from . import utilities
def backgroundFalsePosErrors( coco_analyze, imgs_info, saveDir ):
loc_dir = saveDir + '/background_errors/false_positives'
if not os.path.exists(loc_dir):
os.makedirs(loc_dir)
f = open('%s/std_out.txt'%loc_dir, 'w')
f.write("Running Analysis: [Background False Positives]\n\n")
tic = time.time()
paths = {}
oksThrs = [.5,.55,.6,.65,.7,.75,.8,.85,.9,.95]
areaRngs = [[32**2,1e5**2]]
areaRngLbls = ['all']
coco_analyze.params.areaRng = areaRngs
coco_analyze.params.areaRngLbl = areaRngLbls
coco_analyze.params.oksThrs = oksThrs
coco_analyze.cocoEval.params.useGtIgnore = 0
coco_analyze.cocoEval.params.gtIgnoreIds = []
coco_analyze.analyze(check_kpts=False, check_scores=False, check_bckgd=True)
badFalsePos = coco_analyze.false_pos_dts['all','0.5']
for tind, t in enumerate(coco_analyze.params.oksThrs):
badFalsePos = badFalsePos & coco_analyze.false_pos_dts['all',str(t)]
fp_dts = [d for d in coco_analyze.corrected_dts['all'] if d['id'] in badFalsePos]
f.write("Num. detections: [%d]\n"%len(coco_analyze.corrected_dts['all']))
for oks in oksThrs:
f.write("OKS thresh: [%f]\n"%oks)
f.write(" - Matches: [%d]\n"%len(coco_analyze.bckgd_err_matches[areaRngLbls[0], str(oks), 'dts']))
f.write(" - Bckgd. FP: [%d]\n"%len(coco_analyze.false_pos_dts[areaRngLbls[0],str(oks)]))
sorted_fps = sorted(fp_dts, key=lambda k: -k['score'])
show_fp = sorted_fps[0:4] + sorted_fps[-4:]
f.write("\nBackground False Positive Errors:\n")
for tind, t in enumerate(show_fp):
name = 'bckd_false_pos_%d'%tind
paths[name] = "%s/%s.pdf"%(loc_dir,name)
f.write("Image_id, detection_id, score: [%d][%d][%.3f]\n"%(t['image_id'],t['id'],t['score']))
utilities.show_dets([t],[],imgs_info[t['image_id']],paths[name])
a = [d['score'] for d in coco_analyze.corrected_dts['all']]
p_20 = np.percentile(a, 20); p_40 = np.percentile(a, 40)
p_60 = np.percentile(a, 60); p_80 = np.percentile(a, 80)
f.write("\nPercentiles of the scores of all Detections:\n")
f.write(" - 20th perc. score:[%.3f]; num. dts:[%d]\n"%(p_20,len([d for d in coco_analyze.corrected_dts['all'] if d['score']<=p_20])))
f.write(" - 40th perc. score:[%.3f]; num. dts:[%d]\n"%(p_40,len([d for d in coco_analyze.corrected_dts['all'] if d['score']<=p_40])))
f.write(" - 60th perc. score:[%.3f]; num. dts:[%d]\n"%(p_60,len([d for d in coco_analyze.corrected_dts['all'] if d['score']<=p_60])))
f.write(" - 80th perc. score:[%.3f]; num. dts:[%d]\n"%(p_80,len([d for d in coco_analyze.corrected_dts['all'] if d['score']<=p_80])))
fig, ax = plt.subplots(figsize=(10,10))
ax.set_facecolor('lightgray')
bins = [min(a),p_20,p_40,p_60,p_80,max(a)]
plt.hist([d['score'] for d in sorted_fps],bins=bins,color='b')
plt.xticks(bins,['','20th%','40th%','60th%','80th%',''],rotation='vertical')
plt.grid()
plt.xlabel('All Detection Score Percentiles',fontsize=20)
plt.title('Histogram of False Positive Scores',fontsize=20)
path = "%s/bckd_false_pos_scores_histogram.pdf"%(loc_dir)
paths['false_pos_score_hist'] = path
plt.savefig(path, bbox_inches='tight')
plt.close()
high_score_fp = [d for d in sorted_fps if p_80 <= d['score']]
max_height = max([d['bbox'][3] for d in high_score_fp])
min_height = min([d['bbox'][3] for d in high_score_fp])
max_width = max([d['bbox'][2] for d in high_score_fp])
min_width = min([d['bbox'][2] for d in high_score_fp])
f.write("\nBackground False Positives Bounding Box Dimenstions:\n")
f.write(" - Min width: [%d]\n"%min_width)
f.write(" - Max width: [%d]\n"%max_width)
f.write(" - Min height: [%d]\n"%min_height)
f.write(" - Max height: [%d]\n"%max_height)
ar_pic = np.zeros((int(max_height)+1,int(max_width)+1))
ar_pic_2 = np.zeros((30,30))
ar_bins = list(range(10))+list(range(10,100,10))+list(range(100,1000,100))+[1000]
ar_pic_3 = np.zeros((10,10))
ar_bins_3 = [np.power(2,x) for x in range(11)]
areaRngs = [[0, 32 ** 2],[32 ** 2, 64 ** 2],[64 ** 2, 96 ** 2],[96 ** 2, 128 ** 2],[128 ** 2, 1e5 ** 2]]
areaRngLbls = ['small','medium','large','xlarge','xxlarge']
small = 0; medium = 0; large = 0; xlarge = 0; xxlarge = 0
num_people_ranges = [[0,0],[1,1],[2,4],[5,8],[9,100]]
num_people_labels = ['none','one','small grp.','large grp.', 'crowd']
no_people = 0; one = 0; small_grp = 0; large_grp = 0; crowd = 0
for t in high_score_fp:
t_width = int(t['bbox'][2])
t_height = int(t['bbox'][3])
ar_pic[0:t_height,0:t_width] += 1
if t_width < 1024 and t_height < 1024:
col = [i for i in range(len(ar_bins)-1) if ar_bins[i]<t_width<ar_bins[i+1]]
row = [i for i in range(len(ar_bins)-1) if ar_bins[i]<t_height<ar_bins[i+1]]
ar_pic_2[row,col] += 1
col = [i for i in range(len(ar_bins_3)-1) if ar_bins_3[i]<t_width<ar_bins_3[i+1]]
row = [i for i in range(len(ar_bins_3)-1) if ar_bins_3[i]<t_height<ar_bins_3[i+1]]
ar_pic_3[row,col] += 1
else:
print("False Positive bbox has a side larger than 1024 pixels.")
print("Change lists ar_bins_2 and ar_bins_3 to include larger bins.")
assert(False)
area = t_width * t_height * .5
if areaRngs[0][0] <= area < areaRngs[0][1]:
small += 1
elif areaRngs[1][0] <= area < areaRngs[1][1]:
medium += 1
elif areaRngs[2][0] <= area < areaRngs[2][1]:
large += 1
elif areaRngs[3][0] <= area < areaRngs[3][1]:
xlarge += 1
elif areaRngs[4][0] <= area < areaRngs[4][1]:
xxlarge += 1
anns = coco_analyze.cocoGt.loadAnns(coco_analyze.cocoGt.getAnnIds(t['image_id']))
iscrowd = [ann['iscrowd'] for ann in anns]
num_people = len(anns) if sum(iscrowd)==0 else 100
if num_people_ranges[0][0] <= num_people <= num_people_ranges[0][1]:
no_people += 1
elif num_people_ranges[1][0] <= num_people <= num_people_ranges[1][1]:
one += 1
elif num_people_ranges[2][0] <= num_people <= num_people_ranges[2][1]:
small_grp += 1
elif num_people_ranges[3][0] <= num_people <= num_people_ranges[3][1]:
large_grp += 1
elif num_people_ranges[4][0] <= num_people <= num_people_ranges[4][1]:
crowd += 1
f.write("\nNumber of people in images with Background False Positives:\n")
f.write(" - No people: [%d]\n"%no_people)
f.write(" - One person: [%d]\n"%one)
f.write(" - Small group (2-4): [%d]\n"%small_grp)
f.write(" - Large Group (5-8): [%d]\n"%large_grp)
f.write(" - Crowd (>=9): [%d]\n"%crowd)
f.write("\nArea size (in pixels) of Background False Positives:\n")
f.write(" - Small (%d,%d): [%d]\n"%(areaRngs[0][0],areaRngs[0][1],small))
f.write(" - Medium (%d,%d): [%d]\n"%(areaRngs[1][0],areaRngs[1][1],medium))
f.write(" - Large (%d,%d): [%d]\n"%(areaRngs[2][0],areaRngs[2][1],large))
f.write(" - X-Large (%d,%d): [%d]\n"%(areaRngs[3][0],areaRngs[3][1],xlarge))
f.write(" - XX-Large (%d,%d): [%d]\n"%(areaRngs[4][0],areaRngs[4][1],xxlarge))
plt.figure(figsize=(10,10))
plt.imshow(ar_pic,origin='lower')
plt.colorbar()
plt.title('BBox Aspect Ratio',fontsize=20)
plt.xlabel('Width (px)',fontsize=20)
plt.ylabel('Height (px)',fontsize=20)
path = "%s/bckd_false_pos_bbox_aspect_ratio.pdf"%(loc_dir)
plt.savefig(path, bbox_inches='tight')
plt.close()
fig, ax = plt.subplots(figsize=(10,10))
plt.imshow(ar_pic_2,origin='lower')
plt.xticks(range(1,len(ar_bins)+1),["%d"%(x) for x in ar_bins],rotation='vertical')
plt.yticks(range(1,len(ar_bins)+1),["%d"%(x) for x in ar_bins])
plt.colorbar()
plt.grid()
plt.title('BBox Aspect Ratio',fontsize=20)
plt.xlabel('Width (px)',fontsize=20)
plt.ylabel('Height (px)',fontsize=20)
path = "%s/bckd_false_pos_bbox_aspect_ratio_2.pdf"%(loc_dir)
plt.savefig(path, bbox_inches='tight')
plt.close()
fig, ax = plt.subplots(figsize=(10,10))
plt.imshow(ar_pic_3,origin='lower')
plt.xticks([-.5 + x for x in range(11)],["%d"%(x) for x in ar_bins_3])
plt.yticks([-.5 + x for x in range(11)],["%d"%(x) for x in ar_bins_3])
plt.colorbar()
plt.grid()
plt.title('BBox Aspect Ratio',fontsize=20)
plt.xlabel('Width (px)',fontsize=20)
plt.ylabel('Height (px)',fontsize=20)
path = "%s/bckd_false_pos_bbox_aspect_ratio_3.pdf"%(loc_dir)
paths['false_pos_bbox_ar'] = path
plt.savefig(path, bbox_inches='tight')
plt.close()
fig, ax = plt.subplots(figsize=(10,10))
ax.set_facecolor('lightgray')
plt.bar(range(5),[small,medium,large,xlarge,xxlarge],color='g',align='center')
plt.xticks(range(5),areaRngLbls)
plt.grid()
plt.title('Histogram of Area Size',fontsize=20)
path = "%s/bckd_false_pos_area_histogram.pdf"%(loc_dir)
paths['false_pos_bbox_area_hist'] = path
plt.savefig(path, bbox_inches='tight')
plt.close()
fig, ax = plt.subplots(figsize=(10,10))
ax.set_facecolor('lightgray')
plt.bar(range(5),[no_people,one,small_grp,large_grp,crowd],color='g',align='center')
plt.xticks(range(5),num_people_labels)
plt.grid()
plt.title('Histogram of Num. of People in Images',fontsize=20)
path = "%s/bckd_false_pos_num_people_histogram.pdf"%(loc_dir)
paths['false_pos_num_ppl_hist'] = path
plt.savefig(path, bbox_inches='tight')
plt.close()
f.write("\nDone, (t=%.2fs)."%(time.time()-tic))
f.close()
return paths