-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathpatch_converter_3d.py
245 lines (218 loc) · 11.5 KB
/
patch_converter_3d.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
import numpy as np
from math import cos, sin, radians
import copy
from patch_converter import Patch_converter
import torch
from torchvision.transforms.functional import perspective, InterpolationMode, resize as img_resize
class Patch_converter_3d(object):
def __init__(self, model_name,
angle_range=None, # np.arange(-10, 10 + 1, 1)
dist_range_y=None, # np.arange(5, 10 + 1, 1)
dist_range_x=None, # np.arange(-1, 1 + 0.5, 0.5)
resize_range=None # np.arange(0.5, 1 + 0.1, 0.1)
) -> None:
self.lidar_height = 1.84
self.angle_range = angle_range
self.dist_range_x = dist_range_x
self.dist_range_y = dist_range_y
self.resize_range = resize_range
self.model_name = model_name
self.pc = Patch_converter()
def compose_targeted_patch(self,
ori_patch:torch.Tensor,
ori_mask: torch.Tensor,
ori_patch_area, # Tuple: (top_loc, left_loc, H, W) relative to bevfusion2 size
obj_corners: np.ndarray, # shape: (8, 3), order see: LiDARInstance3DBoxes
transform,
W,
H,
deterministic=True,
rs:np.random.RandomState=None
):
"""torch.Tensor: Coordinates of corners of the target object's bounding box
in shape (8, 3).
Convert the boxes to corners in clockwise order, in form of
``(x0y0z0, x0y0z1, x0y1z1, x0y1z0, x1y0z0, x1y0z1, x1y1z1, x1y1z0)``
.. code-block:: none
up z
front x ^
/ |
/ |
(x1, y0, z1) + ----------- + (x1, y1, z1) ---> p1
/| / |
/ | / |
(x0, y0, z1) + ----------- + + (x1, y1, z0) ---> p2
| / . | /
| / origin | /
left y<-------- + ----------- + (x0, y1, z0) ---> p3
(x0, y0, z0)
"""
if not deterministic:
if self.resize_range is not None:
scale = rs.choice(self.resize_range)
ori_H, ori_W = ori_patch.shape[-2:]
scale_H, scale_W = int(ori_H * scale), int(ori_W * scale)
ori_patch = img_resize(ori_patch, (scale_H, scale_W))
ori_patch = img_resize(ori_patch, (ori_H, ori_W))
back_coords = obj_corners[[2,3,6,7], :]
center_coord = np.sum(back_coords,axis=0) / 4
x_vec = obj_corners[7] - obj_corners[3]
x_vec = x_vec / np.linalg.norm(x_vec) * W/2
z_vec = np.array([0,0,1]) * H/2
p1 = center_coord + x_vec + z_vec
p2 = center_coord + x_vec - z_vec
p3 = center_coord - x_vec - z_vec
p4 = center_coord - x_vec + z_vec
coords = np.hstack([np.vstack([p1, p2, p3, p4]), np.ones((4, 1))])
transform = copy.deepcopy(transform).reshape(4, 4)
coords = coords @ transform.T
coords[:, 2] = np.clip(coords[:, 2], a_min=1e-5, a_max=1e5)
coords[:, 0] /= coords[:, 2]
coords[:, 1] /= coords[:, 2]
coords = coords[:, :2]
# transform of these models are to original image size, so convert to target model input size
# deepint, uvtr, bevformer's transformations are to model input size
if self.model_name == 'bevfusion' or self.model_name == 'bevfusion2' \
or self.model_name == 'transfusion' or self.model_name == 'autoalign':
coords = self.pc.convert_coord_from_ori(target = self.model_name, coords = coords)
coords = coords.astype(np.int32)
coords = coords[[3, 0, 1, 2]] # in the order of tl, tr, br, bl
coords_uv = [[coord[0], coord[1]] for coord in coords]
# from bevfusion scale to target scale
new_patch, new_mask = self.pc.convert_patch(ori_patch, ori_mask, target = self.model_name)
new_patch_area = self.pc.convert_patch_area(ori_patch_area, target = self.model_name, source='bevfusion2')
t, l, h, w = new_patch_area
tl = [l, t]
tr = [l+w, t]
br = [l+w, t+h]
bl = [l, t+h]
patch_bind = torch.cat([new_mask, new_patch], dim=0).clone()
patch_bind = perspective(patch_bind, [tl, tr, br, bl], coords_uv, interpolation=InterpolationMode.BILINEAR, fill=0)
# patch_bind = perspective(patch_bind, [tl, tr, br, bl], coords_uv, interpolation=InterpolationMode.NEAREST, fill=0)
mask_out = patch_bind[[0]]
patch_out = patch_bind[1:4]
mask_out[mask_out.isnan()] = 0
patch_out[patch_out.isnan()] = 0
return patch_out, mask_out
def compose_3d_patch(self,
ori_patch:torch.Tensor,
ori_mask: torch.Tensor,
ori_patch_area, # Tuple: (top_loc, left_loc, H, W) relative to bevfusion2 size
transform,
W,
H,
deterministic=True,
dx=0,
dy=6.7,
alpha=0,
rs:np.random.RandomState=None):
"""
ori_patch: the original patch with optimizable parameters with size relative to bevfusion
ori_mask: the mask used on originial patch
ori_patch_area: the patch area of original patch # Tuple: (top_loc, left_loc, H, W) relative to bevfusion size
"""
if not deterministic:
assert rs is not None
if self.dist_range_x is not None:
dx = rs.choice(self.dist_range_x)
if self.dist_range_y is not None:
dy = rs.choice(self.dist_range_y)
if self.angle_range is not None:
alpha = rs.choice(self.angle_range)
if self.resize_range is not None:
scale = rs.choice(self.resize_range)
ori_H, ori_W = ori_patch.shape[-2:]
scale_H, scale_W = int(ori_H * scale), int(ori_W * scale)
ori_patch = img_resize(ori_patch, (scale_H, scale_W))
ori_patch = img_resize(ori_patch, (ori_H, ori_W))
# print("dx: {}, dy: {}, alpha: {}".format(dx, dy, alpha))
x_vec = np.array((cos(radians(alpha)), sin(radians(alpha))))
y_vec = np.array((-sin(radians(alpha)), cos(radians(alpha))))
m = np.array((dx, dy))
p1 = np.append(m + W/2 * x_vec + H/2 * y_vec, - self.lidar_height)
p2 = np.append(m - W/2 * x_vec + H/2 * y_vec, - self.lidar_height)
p3 = np.append(m - W/2 * x_vec - H/2 * y_vec, - self.lidar_height)
p4 = np.append(m + W/2 * x_vec - H/2 * y_vec, - self.lidar_height)
coords = np.hstack([np.vstack([p1, p2, p3, p4]), np.ones((4, 1))])
transform = copy.deepcopy(transform).reshape(4, 4)
coords = coords @ transform.T
coords[:, 2] = np.clip(coords[:, 2], a_min=1e-5, a_max=1e5)
coords[:, 0] /= coords[:, 2]
coords[:, 1] /= coords[:, 2]
coords = coords[:, :2]
# transform of these models are to original image size, so convert to target model input size
# deepint, uvtr, bevformer's transformations are to model input size
if self.model_name == 'bevfusion' or self.model_name == 'bevfusion2' \
or self.model_name == 'transfusion' or self.model_name == 'autoalign':
coords = self.pc.convert_coord_from_ori(target = self.model_name, coords = coords)
coords = coords.astype(np.int32)
coords = coords[[1, 0, 3, 2]] # in the order of tl, tr, br, bl
coords_uv = [[coord[0], coord[1]] for coord in coords]
# from bevfusion scale to target scale
new_patch, new_mask = self.pc.convert_patch(ori_patch, ori_mask, target = self.model_name)
new_patch_area = self.pc.convert_patch_area(ori_patch_area, target = self.model_name, source='bevfusion2')
t, l, h, w = new_patch_area
tl = [l, t]
tr = [l+w, t]
br = [l+w, t+h]
bl = [l, t+h]
patch_bind = torch.cat([new_mask, new_patch], dim=0).clone()
# patch_bind = perspective(patch_bind, [tl, tr, br, bl], coords_uv, interpolation=InterpolationMode.BILINEAR, fill=0)
patch_bind = perspective(patch_bind, [tl, tr, br, bl], coords_uv, interpolation=InterpolationMode.NEAREST, fill=0)
mask_out = patch_bind[[0]]
patch_out = patch_bind[1:4]
mask_out[mask_out.isnan()] = 0
patch_out[patch_out.isnan()] = 0
return patch_out, mask_out
if __name__ == "__main__":
from model_loader import Model_loader
from my_utils import get_input_data, save_pic
model_loader = Model_loader()
model_name = 'bevfusion2'
model, model_dataloader, target_cfg = model_loader.load_model(model_name)
## test scene-oriented attack
# input_data, data_iter, dataset_idx = get_input_data(model_name, model_dataloader, -1)
# metas = input_data['img_metas'][0].data[0][0]
# k = 1 if model_name == 'bevfusion2' or model_name == 'transfusion' else 0
# transform = metas["lidar2image"][k] if "lidar2image" in metas.keys() else metas["lidar2img"][k]
# pc_3d = Patch_converter_3d(model_name)
# W = 2
# H = 2
# patch_area = (256 - H * 100 - 1, (704 - W * 100) // 2, H * 100, W * 100)
# t, l, h, w = patch_area
# patch = torch.rand((3, 256, 704))
# mask = torch.zeros((1, 256, 704))
# mask[:, t:t+h, l:l+w] = 1
# patch, mask = pc_3d.compose_3d_patch(patch, mask, patch_area, transform, W, H, dx=0, dy=6.7, alpha=0)
# ori_image = input_data['img'][0].data[0][0, k, ...].clone().detach()
# patched_image = ori_image * (1-mask) + patch * mask
# save_pic(patched_image, 0)
## test object-oriented attack
from mmdet3d.core import LiDARInstance3DBoxes
input_data, data_iter, dataset_idx = get_input_data(model_name, model_dataloader, 1542800856950302)
target_idx=17
metas = input_data['img_metas'][0].data[0][0]
k = 1 if model_name == 'bevfusion2' or model_name == 'transfusion' else 0
transform = metas["lidar2image"][k] if "lidar2image" in metas.keys() else metas["lidar2img"][k]
pc_3d = Patch_converter_3d(model_name)
W = 1
H = 1
patch_area = (448 - H * 100 - 1, (800 - W * 100) // 2, H * 100, W * 100)
t, l, h, w = patch_area
patch = torch.rand((3, 448, 800))
mask = torch.zeros((1, 448, 800))
mask[:, t:t+h, l:l+w] = 1
if model_name == "bevfusion" or model_name == 'uvtr':
bboxes = input_data["gt_bboxes_3d"].data[0][0].tensor.numpy()
elif model_name == "deepint" or model_name == 'bevfusion2' or model_name == 'bevformer'\
or model_name == 'transfusion' or model_name == 'autoalign':
bboxes = input_data["gt_bboxes_3d"][0].data[0][0].tensor.numpy()
if model_name == 'bevfusion':
bboxes[..., 2] -= bboxes[..., 5] / 2
bboxes_corners = LiDARInstance3DBoxes(bboxes, box_dim=9).corners.numpy()
obj_corners = bboxes_corners[target_idx]
patch, mask = pc_3d.compose_targeted_patch(patch, mask, patch_area,
obj_corners, transform, W, H)
ori_image = input_data['img'][0].data[0][0, k, ...].clone().detach()
patched_image = ori_image * (1-mask) + patch * mask
save_pic(patched_image, 0)