-
Notifications
You must be signed in to change notification settings - Fork 0
/
smoothing.py
223 lines (194 loc) · 10.6 KB
/
smoothing.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
import numpy as np
import pandas as pd
class Smoothing:
def __init__(self, class_size = 8, window_size=3, confidence_weighting_method=None, bbox_weighting_method=None):
self.window_size = window_size
self.confidence_weighting = self.calculate_weighting(confidence_weighting_method)
self.bbox_weighting = self.calculate_weighting(bbox_weighting_method)
self.right_hand_confidence_history = np.zeros(shape=(window_size, class_size))
self.left_hand_confidence_history = np.zeros(shape=(window_size, class_size))
self.right_hand_bbox_history = np.zeros(shape=(window_size, 4))
self.left_hand_bbox_history = np.zeros(shape=(window_size, 4))
self.mapping_dictionary = dict()
self.last_left_return = {"bbox": np.array([0, 1, 2, 3]), "prediction": "Left_Empty", "confidence": 1}
self.last_right_return = {"bbox": np.array([0, 1, 2, 3]), "prediction": "Right_Empty", "confidence": 1}
self.first_left_prediction = True
self.first_right_prediction = True
self.smoother_params = "window_size" + str(window_size) + "_smoothing" + str(confidence_weighting_method)
def calculate_weighting(self, method=None):
"""
Create weight distribution over window of self.window_size.
Options for linear, log, log+linear (called "superlinear") and no smoothing.
:param method:
:return:
"""
if method == "log":
unnormalized_weights = np.log(range(2, self.window_size + 2))
return unnormalized_weights / np.sum(unnormalized_weights)
elif method == "linear":
unnormalized_weights = np.array(range(2, self.window_size + 2))
return unnormalized_weights / np.sum(unnormalized_weights)
elif method == "exp":
unnormalized_weights = np.exp(range(2, self.window_size + 2))
return unnormalized_weights / np.sum(unnormalized_weights)
elif method == "super-linear":
unnormalized_weights = np.array(range(2, self.window_size + 2)) + np.log(range(2, self.window_size + 2))
return unnormalized_weights / np.sum(unnormalized_weights)
else:
last_frame = np.zeros(self.window_size)
last_frame[-1] = 1.0
return last_frame
def smooth(self, curr_output):
"""
Receives an output from yolo, reduces duplicate predictions (more than one to a given arm)
and returns two predictions one for each arm.
If the object detection model did not make a prediction for a given arm, the last frame is reused.
:param curr_output:
:return:
"""
one_per_arm = self.one_per_arm(curr_output)
if one_per_arm.shape[0] == 2:
return self.smooth_two_outputs(one_per_arm)
elif one_per_arm.shape[0] == 1:
return self.smooth_one_outputs(one_per_arm)
elif one_per_arm.shape[0] == 0:
return self.smooth_zero_outputs(one_per_arm)
else:
print("Failure to assign output to left or right arm")
def smooth_zero_outputs(self, curr_output):
"""
Assuming that both hands really are present, we reuse the last known predictions and bbox of the hands
:return: two dictionaries (left, right) containing bbox, prediction and confidence (these are the keys)
"""
return self.last_left_return, self.last_right_return
def smooth_one_outputs(self, curr_output):
"""
Receive yolo output, smooth and return predictions and smoothed confidence
Infers right or left based on the last known position of each hand
The hand without a prediction is treated as unchanged.
This is a calculated risk based on domain knowledge that both hands are in the video for nearly all time
:param curr_output:
:return: two dictionaries (left, right) containing bbox, prediction and confidence (these are the keys)
"""
row = curr_output.to_numpy()[0, :]
number_label, word_label = row[5], row[6]
self.mapping_dictionary.update({number_label: word_label})
prediction_label = row[6]
hand, _ = prediction_label.split('_', maxsplit=1)
if hand.lower() == "left":
self.update_left_hand(row)
else:
self.update_right_hand(row)
return self.last_left_return, self.last_right_return
def smooth_two_outputs(self, curr_output):
"""
Receive yolo output, smooth and return predictions and smoothed confidence
:param curr_output:
:return: two dictionaries (left, right) with bbox, prediction and confidence (these are the keys)
"""
# associate current output lines to the correct array by x_min
row_1 = curr_output.to_numpy()[0, :]
row_2 = curr_output.to_numpy()[1, :]
# add new labels to dictionary
number_label, word_label = row_1[5], row_1[6]
self.mapping_dictionary.update({number_label: word_label})
number_label, word_label = row_2[5], row_2[6]
self.mapping_dictionary.update({number_label: word_label})
# if row_1[2] < row_2[2]:
# right_hand_row = row_1
# left_hand_row = row_2
# else:
# right_hand_row = row_2
# left_hand_row = row_1
prediction_label = row_1[6]
hand, _ = prediction_label.split('_', maxsplit=1)
if hand.lower() == "left":
right_hand_row = row_2
left_hand_row = row_1
else:
right_hand_row = row_1
left_hand_row = row_2
self.update_left_hand(left_hand_row)
self.update_right_hand(right_hand_row)
return self.last_left_return, self.last_right_return
def update_left_hand(self, row):
"""
Receives a row assigned to the left hand and updates the bounding box and prediction history.
Returns a smoothed bounding box and prediction using the smoothers initalized weighting.
:param row:
:return:
"""
if self.first_left_prediction:
self.first_left_prediction = False
self.left_hand_bbox_history = np.vstack([row[:4].astype(int) for i in range(self.window_size)])
# update confidence history for smoothing
self.left_hand_confidence_history = self.left_hand_confidence_history[1:, :]
new_left_prediction = np.zeros(shape=(1, self.left_hand_confidence_history.shape[1]))
new_left_prediction[0, row[5]] = row[4]
self.left_hand_confidence_history = np.vstack([self.left_hand_confidence_history, new_left_prediction])
left_smoothed_prediction = np.average(self.left_hand_confidence_history, axis=0, weights=self.confidence_weighting)
normalized_left_smoothed_prediction = left_smoothed_prediction / sum(left_smoothed_prediction)
# update bbox history for smoothing
self.left_hand_bbox_history = np.vstack([self.left_hand_bbox_history[1:, :], row[:4].astype(int)])
left_smoothed_bbox = np.average(self.left_hand_bbox_history, axis=0, weights=self.bbox_weighting).astype(int)
# get prediction and confidence
left_prediction = np.argmax(normalized_left_smoothed_prediction)
left_smoothed_confidence = np.max(normalized_left_smoothed_prediction)
self.last_left_return = {"bbox": left_smoothed_bbox, "prediction": self.mapping_dictionary[left_prediction],
"confidence": left_smoothed_confidence}
def update_right_hand(self, row):
"""
Receives a row assigned to the right hand and updates the bounding box and prediction history.
Returns a smoothed bounding box and prediction using the smoothers initalized weighting.
:param row:
:return:
"""
if self.first_right_prediction:
self.first_right_prediction = False
self.right_hand_bbox_history = np.vstack([row[:4].astype(int) for i in range(self.window_size)])
# update confidence history for smoothing
self.right_hand_confidence_history = self.right_hand_confidence_history[1:, :]
new_right_prediction = np.zeros(shape=(1, self.right_hand_confidence_history.shape[1]))
new_right_prediction[0, row[5]] = row[4]
self.right_hand_confidence_history = np.vstack([self.right_hand_confidence_history, new_right_prediction])
right_smoothed_prediction = np.average(self.right_hand_confidence_history, axis=0, weights=self.confidence_weighting)
normalized_right_smoothed_prediction = right_smoothed_prediction / sum(right_smoothed_prediction)
# update bbox history for smoothing
self.right_hand_bbox_history = np.vstack([self.right_hand_bbox_history[1:, :], row[:4].astype(int)])
right_smoothed_bbox = np.average(self.right_hand_bbox_history, axis=0, weights=self.bbox_weighting).astype(int)
# get prediction and confidence
right_prediction = np.argmax(normalized_right_smoothed_prediction)
right_smoothed_confidence = np.max(normalized_right_smoothed_prediction)
self.last_right_return = {"bbox": right_smoothed_bbox, "prediction": self.mapping_dictionary[right_prediction],
"confidence": right_smoothed_confidence}
def one_per_arm(self, curr_output):
"""
Receives the yolo models output and removes duplicate predictions (more than one for each arm)
Returns processed dataframe
:param curr_output:
:return:
"""
left_indices = list()
right_indices = list()
for index, row in curr_output.iterrows():
prediction_label = row[6]
hand, _ = prediction_label.split('_', maxsplit=1)
if hand.lower() == "left":
left_indices.append(index)
else:
right_indices.append(index)
if len(left_indices) != 0:
left_df = curr_output[curr_output.index.isin(left_indices)]
left_max_confidence_df = left_df[left_df.iloc[:, 4] == left_df.iloc[:, 4].max()]
if len(right_indices) != 0:
right_df = curr_output[curr_output.index.isin(right_indices)]
right_max_confidence_df = right_df[right_df.iloc[:, 4] == right_df.iloc[:, 4].max()]
# if there is a prediction for both arms tran use smoothing for two arms
if len(left_indices) != 0 and len(right_indices) != 0:
return pd.concat([left_max_confidence_df, right_max_confidence_df], axis=0)
elif len(left_indices) != 0:
return left_max_confidence_df
elif len(left_indices) != 0:
return right_max_confidence_df
else:
return curr_output