-
Notifications
You must be signed in to change notification settings - Fork 69
/
inference.py
313 lines (249 loc) · 15.5 KB
/
inference.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
311
312
313
"""
EfficientPose (c) by Steinbeis GmbH & Co. KG für Technologietransfer
Haus der Wirtschaft, Willi-Bleicher-Straße 19, 70174 Stuttgart, Germany
Yannick Bukschat: [email protected]
Marcus Vetter: [email protected]
EfficientPose is licensed under a
Creative Commons Attribution-NonCommercial 4.0 International License.
The license can be found in the LICENSE file in the root directory of this source tree
or at http://creativecommons.org/licenses/by-nc/4.0/.
---------------------------------------------------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------
Based on:
Keras EfficientDet implementation (https://github.com/xuannianz/EfficientDet) licensed under the Apache License, Version 2.0
---------------------------------------------------------------------------------------------------------------------------------
The official EfficientDet implementation (https://github.com/google/automl) licensed under the Apache License, Version 2.0
---------------------------------------------------------------------------------------------------------------------------------
EfficientNet Keras implementation (https://github.com/qubvel/efficientnet) licensed under the Apache License, Version 2.0
---------------------------------------------------------------------------------------------------------------------------------
Keras RetinaNet implementation (https://github.com/fizyr/keras-retinanet) licensed under the Apache License, Version 2.0
"""
import cv2
import numpy as np
import os
from tqdm import tqdm
import math
import tensorflow as tf
from model import build_EfficientPose
from utils import preprocess_image
from utils.visualization import draw_detections
def main():
"""
Run EfficientPose in inference mode on all images in a given directory.
"""
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
allow_gpu_growth_memory()
#input parameter
path_to_images = "/Datasets/Linemod_preprocessed/data/02/rgb/"
image_extension = ".png"
phi = 0
path_to_weights = "./weights/phi_0_occlusion_best_ADD(-S).h5"
save_path = "./predictions/occlusion/" #where to save the images or None if the images should be displayed and not saved
# save_path = None
class_to_name = {0: "ape", 1: "can", 2: "cat", 3: "driller", 4: "duck", 5: "eggbox", 6: "glue", 7: "holepuncher"} #Occlusion
#class_to_name = {0: "driller"} #Linemod use a single class with a name of the Linemod objects
score_threshold = 0.5
translation_scale_norm = 1000.0
draw_bbox_2d = False
draw_name = False
#for the linemod and occlusion trained models take this camera matrix and these 3d models. in case you trained a model on a custom dataset you need to take the camera matrix and 3d cuboids from your custom dataset.
camera_matrix = get_linemod_camera_matrix()
name_to_3d_bboxes = get_linemod_3d_bboxes()
class_to_3d_bboxes = {class_idx: name_to_3d_bboxes[name] for class_idx, name in class_to_name.items()}
num_classes = len(class_to_name)
if not os.path.exists(path_to_images):
print("Error: the given path to the images {} does not exist!".format(path_to_images))
return
image_list = [filename for filename in os.listdir(path_to_images) if image_extension in filename]
print("\nInfo: found {} image files".format(len(image_list)))
#build model and load weights
model, image_size = build_model_and_load_weights(phi, num_classes, score_threshold, path_to_weights)
#inferencing
for image_filename in tqdm(image_list):
#load image
image_path = os.path.join(path_to_images, image_filename)
image = cv2.imread(image_path)
original_image = image.copy()
#preprocessing
input_list, scale = preprocess(image, image_size, camera_matrix, translation_scale_norm)
#predict
boxes, scores, labels, rotations, translations = model.predict_on_batch(input_list)
#postprocessing
boxes, scores, labels, rotations, translations = postprocess(boxes, scores, labels, rotations, translations, scale, score_threshold)
draw_detections(original_image,
boxes,
scores,
labels,
rotations,
translations,
class_to_bbox_3D = class_to_3d_bboxes,
camera_matrix = camera_matrix,
label_to_name = class_to_name,
draw_bbox_2d = draw_bbox_2d,
draw_name = draw_name)
if save_path is None:
#display image with predictions
cv2.imshow('image with predictions', original_image)
cv2.waitKey(0)
else:
#only save images to the given path
os.makedirs(save_path, exist_ok = True)
cv2.imwrite(os.path.join(save_path, image_filename.replace(image_extension, "_predicted" + image_extension)), original_image)
def allow_gpu_growth_memory():
"""
Set allow growth GPU memory to true
"""
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
_ = tf.Session(config = config)
def get_linemod_camera_matrix():
"""
Returns:
The Linemod and Occlusion 3x3 camera matrix
"""
return np.array([[572.4114, 0., 325.2611], [0., 573.57043, 242.04899], [0., 0., 1.]], dtype = np.float32)
def get_linemod_3d_bboxes():
"""
Returns:
name_to_3d_bboxes: Dictionary with the Linemod and Occlusion 3D model names as keys and the cuboids as values
"""
name_to_model_info = {"ape": {"diameter": 102.09865663, "min_x": -37.93430000, "min_y": -38.79960000, "min_z": -45.88450000, "size_x": 75.86860000, "size_y": 77.59920000, "size_z": 91.76900000},
"benchvise": {"diameter": 247.50624233, "min_x": -107.83500000, "min_y": -60.92790000, "min_z": -109.70500000, "size_x": 215.67000000, "size_y": 121.85570000, "size_z": 219.41000000},
"cam": {"diameter": 172.49224865, "min_x": -68.32970000, "min_y": -71.51510000, "min_z": -50.24850000, "size_x": 136.65940000, "size_y": 143.03020000, "size_z": 100.49700000},
"can": {"diameter": 201.40358597, "min_x": -50.39580000, "min_y": -90.89790000, "min_z": -96.86700000, "size_x": 100.79160000, "size_y": 181.79580000, "size_z": 193.73400000},
"cat": {"diameter": 154.54551808, "min_x": -33.50540000, "min_y": -63.81650000, "min_z": -58.72830000, "size_x": 67.01070000, "size_y": 127.63300000, "size_z": 117.45660000},
"driller": {"diameter": 261.47178102, "min_x": -114.73800000, "min_y": -37.73570000, "min_z": -104.00100000, "size_x": 229.47600000, "size_y": 75.47140000, "size_z": 208.00200000},
"duck": {"diameter": 108.99920102, "min_x": -52.21460000, "min_y": -38.70380000, "min_z": -42.84850000, "size_x": 104.42920000, "size_y": 77.40760000, "size_z": 85.69700000},
"eggbox": {"diameter": 164.62758848, "min_x": -75.09230000, "min_y": -53.53750000, "min_z": -34.62070000, "size_x": 150.18460000, "size_y": 107.07500000, "size_z": 69.24140000},
"glue": {"diameter": 175.88933422, "min_x": -18.36050000, "min_y": -38.93300000, "min_z": -86.40790000, "size_x": 36.72110000, "size_y": 77.86600000, "size_z": 172.81580000},
"holepuncher": {"diameter": 145.54287471, "min_x": -50.44390000, "min_y": -54.24850000, "min_z": -45.40000000, "size_x": 100.88780000, "size_y": 108.49700000, "size_z": 90.80000000},
"iron": {"diameter": 278.07811733, "min_x": -129.11300000, "min_y": -59.24100000, "min_z": -70.56620000, "size_x": 258.22600000, "size_y": 118.48210000, "size_z": 141.13240000},
"lamp": {"diameter": 282.60129399, "min_x": -101.57300000, "min_y": -58.87630000, "min_z": -106.55800000, "size_x": 203.14600000, "size_y": 117.75250000, "size_z": 213.11600000},
"phone": {"diameter": 212.35825148, "min_x": -46.95910000, "min_y": -73.71670000, "min_z": -92.37370000, "size_x": 93.91810000, "size_y": 147.43340000, "size_z": 184.74740000}}
name_to_3d_bboxes = {name: convert_bbox_3d(model_info) for name, model_info in name_to_model_info.items()}
return name_to_3d_bboxes
def convert_bbox_3d(model_dict):
"""
Converts the 3D model cuboids from the Linemod format (min_x, min_y, min_z, size_x, size_y, size_z) to the (num_corners = 8, num_coordinates = 3) format
Args:
model_dict: Dictionary containing the cuboid information of a single Linemod 3D model in the Linemod format
Returns:
bbox: numpy (8, 3) array containing the 3D model's cuboid, where the first dimension represents the corner points and the second dimension contains the x-, y- and z-coordinates.
"""
#get infos from model dict
min_point_x = model_dict["min_x"]
min_point_y = model_dict["min_y"]
min_point_z = model_dict["min_z"]
size_x = model_dict["size_x"]
size_y = model_dict["size_y"]
size_z = model_dict["size_z"]
bbox = np.zeros(shape = (8, 3))
#lower level
bbox[0, :] = np.array([min_point_x, min_point_y, min_point_z])
bbox[1, :] = np.array([min_point_x + size_x, min_point_y, min_point_z])
bbox[2, :] = np.array([min_point_x + size_x, min_point_y + size_y, min_point_z])
bbox[3, :] = np.array([min_point_x, min_point_y + size_y, min_point_z])
#upper level
bbox[4, :] = np.array([min_point_x, min_point_y, min_point_z + size_z])
bbox[5, :] = np.array([min_point_x + size_x, min_point_y, min_point_z + size_z])
bbox[6, :] = np.array([min_point_x + size_x, min_point_y + size_y, min_point_z + size_z])
bbox[7, :] = np.array([min_point_x, min_point_y + size_y, min_point_z + size_z])
return bbox
def build_model_and_load_weights(phi, num_classes, score_threshold, path_to_weights):
"""
Builds an EfficientPose model and init it with a given weight file
Args:
phi: EfficientPose scaling hyperparameter
num_classes: The number of classes
score_threshold: Minimum score threshold at which a prediction is not filtered out
path_to_weights: Path to the weight file
Returns:
efficientpose_prediction: The EfficientPose model
image_size: Integer image size used as the EfficientPose input resolution for the given phi
"""
print("\nBuilding model...\n")
_, efficientpose_prediction, _ = build_EfficientPose(phi,
num_classes = num_classes,
num_anchors = 9,
freeze_bn = True,
score_threshold = score_threshold,
num_rotation_parameters = 3,
print_architecture = False)
print("\nDone!\n\nLoading weights...")
efficientpose_prediction.load_weights(path_to_weights, by_name=True)
print("Done!")
image_sizes = (512, 640, 768, 896, 1024, 1280, 1408)
image_size = image_sizes[phi]
return efficientpose_prediction, image_size
def preprocess(image, image_size, camera_matrix, translation_scale_norm):
"""
Preprocesses the inputs for EfficientPose
Args:
image: The image to predict
image_size: Input resolution for EfficientPose
camera_matrix: numpy 3x3 array containing the intrinsic camera parameters
translation_scale_norm: factor to change units. EfficientPose internally works with meter and if your dataset unit is mm for example, then you need to set this parameter to 1000
Returns:
input_list: List containing the preprocessed inputs for EfficientPose
scale: The scale factor of the resized input image and the original image
"""
image = image[:, :, ::-1]
image, scale = preprocess_image(image, image_size)
camera_input = get_camera_parameter_input(camera_matrix, scale, translation_scale_norm)
image_batch = np.expand_dims(image, axis=0)
camera_batch = np.expand_dims(camera_input, axis=0)
input_list = [image_batch, camera_batch]
return input_list, scale
def get_camera_parameter_input(camera_matrix, image_scale, translation_scale_norm):
"""
Return the input vector for the camera parameter layer
Args:
camera_matrix: numpy 3x3 array containing the intrinsic camera parameters
image_scale: The scale factor of the resized input image and the original image
translation_scale_norm: factor to change units. EfficientPose internally works with meter and if your dataset unit is mm for example, then you need to set this parameter to 1000
Returns:
input_vector: numpy array [fx, fy, px, py, translation_scale_norm, image_scale]
"""
#input_vector = [fx, fy, px, py, translation_scale_norm, image_scale]
input_vector = np.zeros((6,), dtype = np.float32)
input_vector[0] = camera_matrix[0, 0]
input_vector[1] = camera_matrix[1, 1]
input_vector[2] = camera_matrix[0, 2]
input_vector[3] = camera_matrix[1, 2]
input_vector[4] = translation_scale_norm
input_vector[5] = image_scale
return input_vector
def postprocess(boxes, scores, labels, rotations, translations, scale, score_threshold):
"""
Filter out detections with low confidence scores and rescale the outputs of EfficientPose
Args:
boxes: numpy array [batch_size = 1, max_detections, 4] containing the 2D bounding boxes
scores: numpy array [batch_size = 1, max_detections] containing the confidence scores
labels: numpy array [batch_size = 1, max_detections] containing class label
rotations: numpy array [batch_size = 1, max_detections, 3] containing the axis angle rotation vectors
translations: numpy array [batch_size = 1, max_detections, 3] containing the translation vectors
scale: The scale factor of the resized input image and the original image
score_threshold: Minimum score threshold at which a prediction is not filtered out
Returns:
boxes: numpy array [num_valid_detections, 4] containing the 2D bounding boxes
scores: numpy array [num_valid_detections] containing the confidence scores
labels: numpy array [num_valid_detections] containing class label
rotations: numpy array [num_valid_detections, 3] containing the axis angle rotation vectors
translations: numpy array [num_valid_detections, 3] containing the translation vectors
"""
boxes, scores, labels, rotations, translations = np.squeeze(boxes), np.squeeze(scores), np.squeeze(labels), np.squeeze(rotations), np.squeeze(translations)
# correct boxes for image scale
boxes /= scale
#rescale rotations
rotations *= math.pi
#filter out detections with low scores
indices = np.where(scores[:] > score_threshold)
# select detections
scores = scores[indices]
boxes = boxes[indices]
rotations = rotations[indices]
translations = translations[indices]
labels = labels[indices]
return boxes, scores, labels, rotations, translations
if __name__ == '__main__':
main()