-
Notifications
You must be signed in to change notification settings - Fork 175
/
make_densebox_target.py
213 lines (185 loc) · 8.26 KB
/
make_densebox_target.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
# encoding: utf-8
import os
from typing import Dict, Tuple
import numpy as np
import torch
DUMP_FLAG = False # dump intermediate results for debugging
DUMP_DIR = "dump"
DUMP_SUFFIX = "v2_1"
if not os.path.exists(DUMP_DIR):
os.makedirs(DUMP_DIR)
def make_densebox_target(gt_boxes: np.array, config: Dict) -> Tuple:
""" v2.1,
fix indexing type cast (compatible with previous)
Model training target generation function for densebox
Target processing code changed from numpy to pytorch
Only one resolution layer is taken into consideration
Refined & documented in detail, comparing to precedented version
About Training Accuracy w.r.t. previous version (torch==1.5.1)
siamfcpp-alexnet: ao@got10k-val = 73.3
siamfcpp-googlenet: ao@got10k-val = 76.3
About alignmenet w.r.t. v2
- classification target: aligned
- centerness target: aligned
- bbox target: aligned
Arguments
---------
gt_boxes : np.array
ground truth bounding boxes with class, shape=(N, 5), order=(x0, y0, x1, y1, class)
config: configuration of target making (old format)
Keys
----
x_size : int
search image size
score_size : int
score feature map size
total_stride : int
total stride of backbone
score_offset : int
offset between the edge of score map and the border of the search image
Returns
-------
Tuple
cls_res_final : np.array
class
shape=(N, 1)
ctr_res_final : np.array
shape=(N, 1)
gt_boxes_res_final : np.array
shape=(N, 4)
# previous format
# shape=(N, 6), order=(class, center-ness, left_offset, top_offset, right_offset, bottom_offset)
"""
x_size = config["x_size"]
score_size = config["score_size"]
total_stride = config["total_stride"]
score_offset = config["score_offset"]
eps = 1e-5
raw_height, raw_width = x_size, x_size
# append class dimension to gt_boxes if ignored
if gt_boxes.shape[1] == 4:
gt_boxes = np.concatenate(
[gt_boxes, np.ones(
(gt_boxes.shape[0], 1))], axis=1) # boxes_cnt x 5
gt_boxes = torch.from_numpy(gt_boxes).type(torch.float32)
# gt box area
# TODO: consider change to max - min + 1?
# (#boxes, 4-d_box + 1-d_cls)
# append dummy box (0, 0, 0, 0) at first for convenient
# #boxes++
gt_boxes = torch.cat([torch.zeros(1, 5, dtype=torch.float32), gt_boxes],
dim=0)
gt_boxes_area = (torch.abs(
(gt_boxes[:, 2] - gt_boxes[:, 0]) * (gt_boxes[:, 3] - gt_boxes[:, 1])))
# sort gt_boxes by area, ascending order
# small box priviledged to large box
gt_boxes = gt_boxes[torch.argsort(gt_boxes_area)]
# #boxes
boxes_cnt = len(gt_boxes)
# coordinate meshgrid on image, shape=(H. W)
x_coords = torch.arange(0, raw_width, dtype=torch.int64) # (W, )
y_coords = torch.arange(0, raw_height, dtype=torch.int64) # (H, )
y_coords, x_coords = torch.meshgrid(x_coords, y_coords) # (H, W)
off_l = (x_coords[:, :, np.newaxis, np.newaxis].type(torch.float32) -
gt_boxes[np.newaxis, np.newaxis, :, 0, np.newaxis])
off_t = (y_coords[:, :, np.newaxis, np.newaxis].type(torch.float32) -
gt_boxes[np.newaxis, np.newaxis, :, 1, np.newaxis])
off_r = -(x_coords[:, :, np.newaxis, np.newaxis].type(torch.float32) -
gt_boxes[np.newaxis, np.newaxis, :, 2, np.newaxis])
off_b = -(y_coords[:, :, np.newaxis, np.newaxis].type(torch.float32) -
gt_boxes[np.newaxis, np.newaxis, :, 3, np.newaxis])
if DUMP_FLAG:
off_l.numpy().dump("{}/off_l_{}.npz".format(DUMP_DIR, DUMP_SUFFIX))
off_t.numpy().dump("{}/off_t_{}.npz".format(DUMP_DIR, DUMP_SUFFIX))
off_r.numpy().dump("{}/off_r_{}.npz".format(DUMP_DIR, DUMP_SUFFIX))
off_b.numpy().dump("{}/off_b_{}.npz".format(DUMP_DIR, DUMP_SUFFIX))
# centerness
# (H, W, #boxes, 1-d_centerness)
# CAUTION: division / broadcast operation can vary across computing framework (pytorch/numpy/etc.)
# which may cause computation result misalignement (but should be really slight)
center = ((torch.min(off_l, off_r) * torch.min(off_t, off_b)) /
(torch.max(off_l, off_r) * torch.max(off_t, off_b) + eps))
# TODO: consider using clamp rather than adding epsilon?
# center = ((torch.min(off_l, off_r) * torch.min(off_t, off_b)) /
# torch.clamp(torch.max(off_l, off_r) * torch.max(off_t, off_b), min=eps))
if DUMP_FLAG:
center.numpy().dump("{}/center_{}.npz".format(DUMP_DIR, DUMP_SUFFIX))
# (H, W, #boxes, )
center = torch.squeeze(torch.sqrt(torch.abs(center)), dim=3)
center[:, :, 0] = 0 # mask centerness for dummy box as zero
# (H, W, #boxes, 4)
offset = torch.cat([off_l, off_t, off_r, off_b], dim=3)
if DUMP_FLAG:
offset.numpy().dump("{}/offset_{}.npz".format(DUMP_DIR, DUMP_SUFFIX))
# (#boxes, )
# store cls index of each box
# class 0 is background
# dummy box assigned as 0
cls = gt_boxes[:, 4]
fm_height, fm_width = score_size, score_size # h, w
fm_offset = score_offset
stride = total_stride
# coordinate meshgrid on feature map, shape=(h, w)
x_coords_on_fm = torch.arange(0, fm_width, dtype=torch.int64) # (w, )
y_coords_on_fm = torch.arange(0, fm_height, dtype=torch.int64) # (h, )
y_coords_on_fm, x_coords_on_fm = torch.meshgrid(x_coords_on_fm,
y_coords_on_fm) # (h, w)
y_coords_on_fm = y_coords_on_fm.reshape(-1) # (hxw, ), flattened
x_coords_on_fm = x_coords_on_fm.reshape(-1) # (hxw, ), flattened
# (hxw, #boxes, 4-d_offset_(l/t/r/b), )
offset_on_fm = offset[fm_offset + y_coords_on_fm * stride, fm_offset +
x_coords_on_fm * stride] # will reduce dim by 1
# (hxw, #gt_boxes, )
is_in_boxes = (offset_on_fm > 0).all(dim=2).type(torch.uint8)
# (h, w, #gt_boxes, ), boolean
# valid mask
offset_valid = torch.zeros((fm_height, fm_width, boxes_cnt),
dtype=torch.uint8)
offset_valid[
y_coords_on_fm,
x_coords_on_fm, :] = is_in_boxes #& is_in_layer # xy[:, 0], xy[:, 1] reduce dim by 1 to match is_in_boxes.shape & is_in_layer.shape
offset_valid[:, :, 0] = 0 # h x w x boxes_cnt
# (h, w), boolean
# index of pixel on feature map
# used for indexing on gt_boxes, cls
# if not match any box, fall on dummy box at index 0
# if conflict, choose box with smaller index
# P.S. boxes already ordered by box's area
# Attention: be aware of definition of _argmax_ here
# which is assumed to find the FIRST OCCURENCE of the max value
# currently torch.argmax's behavior is not aligned with np.argmax
# c.f. https://github.com/pytorch/pytorch/issues/22853
hit_gt_ind = np.argmax(offset_valid, axis=2)
# (h, w, 4-d_box)
# gt_boxes
gt_boxes_res = torch.zeros((fm_height, fm_width, 4))
gt_boxes_res[y_coords_on_fm, x_coords_on_fm] = gt_boxes[
hit_gt_ind[y_coords_on_fm, x_coords_on_fm], :4] # gt_boxes: (#boxes, 5)
gt_boxes_res = gt_boxes_res.reshape(-1, 4)
# gt_boxes_res_list.append(gt_boxes_res.reshape(-1, 4))
# (h, w, 1-d_cls_score)
cls_res = torch.zeros((fm_height, fm_width))
cls_res[y_coords_on_fm, x_coords_on_fm] = cls[
hit_gt_ind[y_coords_on_fm, x_coords_on_fm]]
cls_res = cls_res.reshape(-1, 1)
# (h, w, 1-d_centerness)
center_res = torch.zeros((fm_height, fm_width))
center_res[y_coords_on_fm, x_coords_on_fm] = center[
fm_offset + y_coords_on_fm * stride, fm_offset +
x_coords_on_fm * stride, hit_gt_ind[y_coords_on_fm, x_coords_on_fm]]
center_res = center_res.reshape(-1, 1)
return cls_res, center_res, gt_boxes_res
if __name__ == '__main__':
# gt_boxes
gt_boxes = np.asarray([[13, 25, 100, 140, 1]])
config_dict = dict(
x_size=303,
score_size=17,
total_stride=8,
score_offset=(303 - 1 - (17 - 1) * 8) // 2,
)
target = make_densebox_target(gt_boxes, config_dict)
for v in target:
print("{}".format(v.shape))
from IPython import embed
embed()