-
Notifications
You must be signed in to change notification settings - Fork 232
/
j_and_f.py
310 lines (259 loc) · 12.7 KB
/
j_and_f.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
import numpy as np
import math
from scipy.optimize import linear_sum_assignment
from ..utils import TrackEvalException
from ._base_metric import _BaseMetric
from .. import _timing
class JAndF(_BaseMetric):
"""Class which implements the J&F metrics"""
def __init__(self, config=None):
super().__init__()
self.integer_fields = ['num_gt_tracks']
self.float_fields = ['J-Mean', 'J-Recall', 'J-Decay', 'F-Mean', 'F-Recall', 'F-Decay', 'J&F']
self.fields = self.float_fields + self.integer_fields
self.summary_fields = self.float_fields
self.optim_type = 'J' # possible values J, J&F
@_timing.time
def eval_sequence(self, data):
"""Returns J&F metrics for one sequence"""
# Only loaded when run to reduce minimum requirements
from pycocotools import mask as mask_utils
num_timesteps = data['num_timesteps']
num_tracker_ids = data['num_tracker_ids']
num_gt_ids = data['num_gt_ids']
gt_dets = data['gt_dets']
tracker_dets = data['tracker_dets']
gt_ids = data['gt_ids']
tracker_ids = data['tracker_ids']
# get shape of frames
frame_shape = None
if num_gt_ids > 0:
for t in range(num_timesteps):
if len(gt_ids[t]) > 0:
frame_shape = gt_dets[t][0]['size']
break
elif num_tracker_ids > 0:
for t in range(num_timesteps):
if len(tracker_ids[t]) > 0:
frame_shape = tracker_dets[t][0]['size']
break
if frame_shape:
# append all zero masks for timesteps in which tracks do not have a detection
zero_padding = np.zeros((frame_shape), order= 'F').astype(np.uint8)
padding_mask = mask_utils.encode(zero_padding)
for t in range(num_timesteps):
gt_id_det_mapping = {gt_ids[t][i]: gt_dets[t][i] for i in range(len(gt_ids[t]))}
gt_dets[t] = [gt_id_det_mapping[index] if index in gt_ids[t] else padding_mask for index
in range(num_gt_ids)]
tracker_id_det_mapping = {tracker_ids[t][i]: tracker_dets[t][i] for i in range(len(tracker_ids[t]))}
tracker_dets[t] = [tracker_id_det_mapping[index] if index in tracker_ids[t] else padding_mask for index
in range(num_tracker_ids)]
# also perform zero padding if number of tracker IDs < number of ground truth IDs
if num_tracker_ids < num_gt_ids:
diff = num_gt_ids - num_tracker_ids
for t in range(num_timesteps):
tracker_dets[t] = tracker_dets[t] + [padding_mask for _ in range(diff)]
num_tracker_ids += diff
j = self._compute_j(gt_dets, tracker_dets, num_gt_ids, num_tracker_ids, num_timesteps)
# boundary threshold for F computation
bound_th = 0.008
# perform matching
if self.optim_type == 'J&F':
f = np.zeros_like(j)
for k in range(num_tracker_ids):
for i in range(num_gt_ids):
f[k, i, :] = self._compute_f(gt_dets, tracker_dets, k, i, bound_th)
optim_metrics = (np.mean(j, axis=2) + np.mean(f, axis=2)) / 2
row_ind, col_ind = linear_sum_assignment(- optim_metrics)
j_m = j[row_ind, col_ind, :]
f_m = f[row_ind, col_ind, :]
elif self.optim_type == 'J':
optim_metrics = np.mean(j, axis=2)
row_ind, col_ind = linear_sum_assignment(- optim_metrics)
j_m = j[row_ind, col_ind, :]
f_m = np.zeros_like(j_m)
for i, (tr_ind, gt_ind) in enumerate(zip(row_ind, col_ind)):
f_m[i] = self._compute_f(gt_dets, tracker_dets, tr_ind, gt_ind, bound_th)
else:
raise TrackEvalException('Unsupported optimization type %s for J&F metric.' % self.optim_type)
# append zeros for false negatives
if j_m.shape[0] < data['num_gt_ids']:
diff = data['num_gt_ids'] - j_m.shape[0]
j_m = np.concatenate((j_m, np.zeros((diff, j_m.shape[1]))), axis=0)
f_m = np.concatenate((f_m, np.zeros((diff, f_m.shape[1]))), axis=0)
# compute the metrics for each ground truth track
res = {
'J-Mean': [np.nanmean(j_m[i, :]) for i in range(j_m.shape[0])],
'J-Recall': [np.nanmean(j_m[i, :] > 0.5 + np.finfo('float').eps) for i in range(j_m.shape[0])],
'F-Mean': [np.nanmean(f_m[i, :]) for i in range(f_m.shape[0])],
'F-Recall': [np.nanmean(f_m[i, :] > 0.5 + np.finfo('float').eps) for i in range(f_m.shape[0])],
'J-Decay': [],
'F-Decay': []
}
n_bins = 4
ids = np.round(np.linspace(1, data['num_timesteps'], n_bins + 1) + 1e-10) - 1
ids = ids.astype(np.uint8)
for k in range(j_m.shape[0]):
d_bins_j = [j_m[k][ids[i]:ids[i + 1] + 1] for i in range(0, n_bins)]
res['J-Decay'].append(np.nanmean(d_bins_j[0]) - np.nanmean(d_bins_j[3]))
for k in range(f_m.shape[0]):
d_bins_f = [f_m[k][ids[i]:ids[i + 1] + 1] for i in range(0, n_bins)]
res['F-Decay'].append(np.nanmean(d_bins_f[0]) - np.nanmean(d_bins_f[3]))
# count number of tracks for weighting of the result
res['num_gt_tracks'] = len(res['J-Mean'])
for field in ['J-Mean', 'J-Recall', 'J-Decay', 'F-Mean', 'F-Recall', 'F-Decay']:
res[field] = np.mean(res[field])
res['J&F'] = (res['J-Mean'] + res['F-Mean']) / 2
return res
def combine_sequences(self, all_res):
"""Combines metrics across all sequences"""
res = {'num_gt_tracks': self._combine_sum(all_res, 'num_gt_tracks')}
for field in self.summary_fields:
res[field] = self._combine_weighted_av(all_res, field, res, weight_field='num_gt_tracks')
return res
def combine_classes_class_averaged(self, all_res, ignore_empty_classes=False):
"""Combines metrics across all classes by averaging over the class values
'ignore empty classes' is not yet implemented here.
"""
res = {'num_gt_tracks': self._combine_sum(all_res, 'num_gt_tracks')}
for field in self.float_fields:
res[field] = np.mean([v[field] for v in all_res.values()])
return res
def combine_classes_det_averaged(self, all_res):
"""Combines metrics across all classes by averaging over the detection values"""
res = {'num_gt_tracks': self._combine_sum(all_res, 'num_gt_tracks')}
for field in self.float_fields:
res[field] = np.mean([v[field] for v in all_res.values()])
return res
@staticmethod
def _seg2bmap(seg, width=None, height=None):
"""
From a segmentation, compute a binary boundary map with 1 pixel wide
boundaries. The boundary pixels are offset by 1/2 pixel towards the
origin from the actual segment boundary.
Arguments:
seg : Segments labeled from 1..k.
width : Width of desired bmap <= seg.shape[1]
height : Height of desired bmap <= seg.shape[0]
Returns:
bmap (ndarray): Binary boundary map.
David Martin <dmartin@eecs.berkeley.edu>
January 2003
"""
seg = seg.astype(np.bool)
seg[seg > 0] = 1
assert np.atleast_3d(seg).shape[2] == 1
width = seg.shape[1] if width is None else width
height = seg.shape[0] if height is None else height
h, w = seg.shape[:2]
ar1 = float(width) / float(height)
ar2 = float(w) / float(h)
assert not (
width > w | height > h | abs(ar1 - ar2) > 0.01
), "Can" "t convert %dx%d seg to %dx%d bmap." % (w, h, width, height)
e = np.zeros_like(seg)
s = np.zeros_like(seg)
se = np.zeros_like(seg)
e[:, :-1] = seg[:, 1:]
s[:-1, :] = seg[1:, :]
se[:-1, :-1] = seg[1:, 1:]
b = seg ^ e | seg ^ s | seg ^ se
b[-1, :] = seg[-1, :] ^ e[-1, :]
b[:, -1] = seg[:, -1] ^ s[:, -1]
b[-1, -1] = 0
if w == width and h == height:
bmap = b
else:
bmap = np.zeros((height, width))
for x in range(w):
for y in range(h):
if b[y, x]:
j = 1 + math.floor((y - 1) + height / h)
i = 1 + math.floor((x - 1) + width / h)
bmap[j, i] = 1
return bmap
@staticmethod
def _compute_f(gt_data, tracker_data, tracker_data_id, gt_id, bound_th):
"""
Perform F computation for a given gt and a given tracker ID. Adapted from
https://github.com/davisvideochallenge/davis2017-evaluation
:param gt_data: the encoded gt masks
:param tracker_data: the encoded tracker masks
:param tracker_data_id: the tracker ID
:param gt_id: the ground truth ID
:param bound_th: boundary threshold parameter
:return: the F value for the given tracker and gt ID
"""
# Only loaded when run to reduce minimum requirements
from pycocotools import mask as mask_utils
from skimage.morphology import disk
import cv2
f = np.zeros(len(gt_data))
for t, (gt_masks, tracker_masks) in enumerate(zip(gt_data, tracker_data)):
curr_tracker_mask = mask_utils.decode(tracker_masks[tracker_data_id])
curr_gt_mask = mask_utils.decode(gt_masks[gt_id])
bound_pix = bound_th if bound_th >= 1 - np.finfo('float').eps else \
np.ceil(bound_th * np.linalg.norm(curr_tracker_mask.shape))
# Get the pixel boundaries of both masks
fg_boundary = JAndF._seg2bmap(curr_tracker_mask)
gt_boundary = JAndF._seg2bmap(curr_gt_mask)
# fg_dil = binary_dilation(fg_boundary, disk(bound_pix))
fg_dil = cv2.dilate(fg_boundary.astype(np.uint8), disk(bound_pix).astype(np.uint8))
# gt_dil = binary_dilation(gt_boundary, disk(bound_pix))
gt_dil = cv2.dilate(gt_boundary.astype(np.uint8), disk(bound_pix).astype(np.uint8))
# Get the intersection
gt_match = gt_boundary * fg_dil
fg_match = fg_boundary * gt_dil
# Area of the intersection
n_fg = np.sum(fg_boundary)
n_gt = np.sum(gt_boundary)
# % Compute precision and recall
if n_fg == 0 and n_gt > 0:
precision = 1
recall = 0
elif n_fg > 0 and n_gt == 0:
precision = 0
recall = 1
elif n_fg == 0 and n_gt == 0:
precision = 1
recall = 1
else:
precision = np.sum(fg_match) / float(n_fg)
recall = np.sum(gt_match) / float(n_gt)
# Compute F measure
if precision + recall == 0:
f_val = 0
else:
f_val = 2 * precision * recall / (precision + recall)
f[t] = f_val
return f
@staticmethod
def _compute_j(gt_data, tracker_data, num_gt_ids, num_tracker_ids, num_timesteps):
"""
Computation of J value for all ground truth IDs and all tracker IDs in the given sequence. Adapted from
https://github.com/davisvideochallenge/davis2017-evaluation
:param gt_data: the ground truth masks
:param tracker_data: the tracker masks
:param num_gt_ids: the number of ground truth IDs
:param num_tracker_ids: the number of tracker IDs
:param num_timesteps: the number of timesteps
:return: the J values
"""
# Only loaded when run to reduce minimum requirements
from pycocotools import mask as mask_utils
j = np.zeros((num_tracker_ids, num_gt_ids, num_timesteps))
for t, (time_gt, time_data) in enumerate(zip(gt_data, tracker_data)):
# run length encoded masks with pycocotools
area_gt = mask_utils.area(time_gt)
time_data = list(time_data)
area_tr = mask_utils.area(time_data)
area_tr = np.repeat(area_tr[:, np.newaxis], len(area_gt), axis=1)
area_gt = np.repeat(area_gt[np.newaxis, :], len(area_tr), axis=0)
# mask iou computation with pycocotools
ious = np.atleast_2d(mask_utils.iou(time_data, time_gt, [0]*len(time_gt)))
# set iou to 1 if both masks are close to 0 (no ground truth and no predicted mask in timestep)
ious[np.isclose(area_tr, 0) & np.isclose(area_gt, 0)] = 1
assert (ious >= 0 - np.finfo('float').eps).all()
assert (ious <= 1 + np.finfo('float').eps).all()
j[..., t] = ious
return j