File size: 10,786 Bytes
4121bec
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch

from detectron2.structures import Boxes, RotatedBoxes, pairwise_iou, pairwise_iou_rotated

""" Soft-NMS Pull request from: https://github.com/facebookresearch/detectron2/pull/1183/files
"""

def soft_nms(boxes, scores, method, gaussian_sigma, linear_threshold, prune_threshold):
    """
    Performs soft non-maximum suppression algorithm on axis aligned boxes
    Args:
        boxes (Tensor[N, 5]):
           boxes where NMS will be performed. They
           are expected to be in (x_ctr, y_ctr, width, height, angle_degrees) format
        scores (Tensor[N]):
           scores for each one of the boxes
        method (str):
           one of ['gaussian', 'linear', 'hard']
           see paper for details. users encouraged not to use "hard", as this is the
           same nms available elsewhere in detectron2
        gaussian_sigma (float):
           parameter for Gaussian penalty function
        linear_threshold (float):
           iou threshold for applying linear decay. Nt from the paper
           re-used as threshold for standard "hard" nms
        prune_threshold (float):
           boxes with scores below this threshold are pruned at each iteration.
           Dramatically reduces computation time. Authors use values in [10e-4, 10e-2]
    Returns:
        tuple(Tensor, Tensor):
            [0]: int64 tensor with the indices of the elements that have been kept
            by Soft NMS, sorted in decreasing order of scores
            [1]: float tensor with the re-scored scores of the elements that were kept
"""
    return _soft_nms(
        Boxes,
        pairwise_iou,
        boxes,
        scores,
        method,
        gaussian_sigma,
        linear_threshold,
        prune_threshold,
    )


def soft_nms_rotated(boxes, scores, method, gaussian_sigma, linear_threshold, prune_threshold):
    """
    Performs soft non-maximum suppression algorithm on rotated boxes
    Args:
        boxes (Tensor[N, 5]):
           boxes where NMS will be performed. They
           are expected to be in (x_ctr, y_ctr, width, height, angle_degrees) format
        scores (Tensor[N]):
           scores for each one of the boxes
        method (str):
           one of ['gaussian', 'linear', 'hard']
           see paper for details. users encouraged not to use "hard", as this is the
           same nms available elsewhere in detectron2
        gaussian_sigma (float):
           parameter for Gaussian penalty function
        linear_threshold (float):
           iou threshold for applying linear decay. Nt from the paper
           re-used as threshold for standard "hard" nms
        prune_threshold (float):
           boxes with scores below this threshold are pruned at each iteration.
           Dramatically reduces computation time. Authors use values in [10e-4, 10e-2]
    Returns:
        tuple(Tensor, Tensor):
            [0]: int64 tensor with the indices of the elements that have been kept
            by Soft NMS, sorted in decreasing order of scores
            [1]: float tensor with the re-scored scores of the elements that were kept    """
    return _soft_nms(
        RotatedBoxes,
        pairwise_iou_rotated,
        boxes,
        scores,
        method,
        gaussian_sigma,
        linear_threshold,
        prune_threshold,
    )


def batched_soft_nms(
    boxes, scores, idxs, method, gaussian_sigma, linear_threshold, prune_threshold
):
    """
    Performs soft non-maximum suppression in a batched fashion.
    Each index value correspond to a category, and NMS
    will not be applied between elements of different categories.
    Args:
        boxes (Tensor[N, 4]):
           boxes where NMS will be performed. They
           are expected to be in (x1, y1, x2, y2) format
        scores (Tensor[N]):
           scores for each one of the boxes
        idxs (Tensor[N]):
           indices of the categories for each one of the boxes.
        method (str):
           one of ['gaussian', 'linear', 'hard']
           see paper for details. users encouraged not to use "hard", as this is the
           same nms available elsewhere in detectron2
        gaussian_sigma (float):
           parameter for Gaussian penalty function
        linear_threshold (float):
           iou threshold for applying linear decay. Nt from the paper
           re-used as threshold for standard "hard" nms
        prune_threshold (float):
           boxes with scores below this threshold are pruned at each iteration.
           Dramatically reduces computation time. Authors use values in [10e-4, 10e-2]
    Returns:
        tuple(Tensor, Tensor):
            [0]: int64 tensor with the indices of the elements that have been kept
            by Soft NMS, sorted in decreasing order of scores
            [1]: float tensor with the re-scored scores of the elements that were kept
    """
    if boxes.numel() == 0:
        return (
            torch.empty((0,), dtype=torch.int64, device=boxes.device),
            torch.empty((0,), dtype=torch.float32, device=scores.device),
        )
    # strategy: in order to perform NMS independently per class.
    # we add an offset to all the boxes. The offset is dependent
    # only on the class idx, and is large enough so that boxes
    # from different classes do not overlap
    max_coordinate = boxes.max()
    offsets = idxs.to(boxes) * (max_coordinate + 1)
    boxes_for_nms = boxes + offsets[:, None]
    return soft_nms(
        boxes_for_nms, scores, method, gaussian_sigma, linear_threshold, prune_threshold
    )


def batched_soft_nms_rotated(
    boxes, scores, idxs, method, gaussian_sigma, linear_threshold, prune_threshold
):
    """
    Performs soft non-maximum suppression in a batched fashion on rotated bounding boxes.
    Each index value correspond to a category, and NMS
    will not be applied between elements of different categories.
    Args:
        boxes (Tensor[N, 5]):
           boxes where NMS will be performed. They
           are expected to be in (x_ctr, y_ctr, width, height, angle_degrees) format
        scores (Tensor[N]):
           scores for each one of the boxes
        idxs (Tensor[N]):
           indices of the categories for each one of the boxes.
        method (str):
           one of ['gaussian', 'linear', 'hard']
           see paper for details. users encouraged not to use "hard", as this is the
           same nms available elsewhere in detectron2
        gaussian_sigma (float):
           parameter for Gaussian penalty function
        linear_threshold (float):
           iou threshold for applying linear decay. Nt from the paper
           re-used as threshold for standard "hard" nms
        prune_threshold (float):
           boxes with scores below this threshold are pruned at each iteration.
           Dramatically reduces computation time. Authors use values in [10e-4, 10e-2]
    Returns:
        tuple(Tensor, Tensor):
            [0]: int64 tensor with the indices of the elements that have been kept
            by Soft NMS, sorted in decreasing order of scores
            [1]: float tensor with the re-scored scores of the elements that were kept
    """
    if boxes.numel() == 0:
        return (
            torch.empty((0,), dtype=torch.int64, device=boxes.device),
            torch.empty((0,), dtype=torch.float32, device=scores.device),
        )
    # strategy: in order to perform NMS independently per class.
    # we add an offset to all the boxes. The offset is dependent
    # only on the class idx, and is large enough so that boxes
    # from different classes do not overlap
    max_coordinate = boxes[:, :2].max() + torch.norm(boxes[:, 2:4], 2, dim=1).max()
    offsets = idxs.to(boxes) * (max_coordinate + 1)
    boxes_for_nms = boxes.clone()
    boxes_for_nms[:, :2] += offsets[:, None]
    return soft_nms_rotated(
        boxes_for_nms, scores, method, gaussian_sigma, linear_threshold, prune_threshold
    )


def _soft_nms(
    box_class,
    pairwise_iou_func,
    boxes,
    scores,
    method,
    gaussian_sigma,
    linear_threshold,
    prune_threshold,
):
    """
    Soft non-max suppression algorithm.
    Implementation of [Soft-NMS -- Improving Object Detection With One Line of Codec]
    (https://arxiv.org/abs/1704.04503)
    Args:
        box_class (cls): one of Box, RotatedBoxes
        pairwise_iou_func (func): one of pairwise_iou, pairwise_iou_rotated
        boxes (Tensor[N, ?]):
           boxes where NMS will be performed
           if Boxes, in (x1, y1, x2, y2) format
           if RotatedBoxes, in (x_ctr, y_ctr, width, height, angle_degrees) format
        scores (Tensor[N]):
           scores for each one of the boxes
        method (str):
           one of ['gaussian', 'linear', 'hard']
           see paper for details. users encouraged not to use "hard", as this is the
           same nms available elsewhere in detectron2
        gaussian_sigma (float):
           parameter for Gaussian penalty function
        linear_threshold (float):
           iou threshold for applying linear decay. Nt from the paper
           re-used as threshold for standard "hard" nms
        prune_threshold (float):
           boxes with scores below this threshold are pruned at each iteration.
           Dramatically reduces computation time. Authors use values in [10e-4, 10e-2]
    Returns:
        tuple(Tensor, Tensor):
            [0]: int64 tensor with the indices of the elements that have been kept
            by Soft NMS, sorted in decreasing order of scores
            [1]: float tensor with the re-scored scores of the elements that were kept
    """
    boxes = boxes.clone()
    scores = scores.clone()
    idxs = torch.arange(scores.size()[0])

    idxs_out = []
    scores_out = []

    while scores.numel() > 0:
        top_idx = torch.argmax(scores)
        idxs_out.append(idxs[top_idx].item())
        scores_out.append(scores[top_idx].item())

        top_box = boxes[top_idx]
        ious = pairwise_iou_func(box_class(top_box.unsqueeze(0)), box_class(boxes))[0]

        if method == "linear":
            decay = torch.ones_like(ious)
            decay_mask = ious > linear_threshold
            decay[decay_mask] = 1 - ious[decay_mask]
        elif method == "gaussian":
            decay = torch.exp(-torch.pow(ious, 2) / gaussian_sigma)
        elif method == "hard":  # standard NMS
            decay = (ious < linear_threshold).float()
        else:
            raise NotImplementedError("{} soft nms method not implemented.".format(method))

        scores *= decay
        keep = scores > prune_threshold
        keep[top_idx] = False

        boxes = boxes[keep]
        scores = scores[keep]
        idxs = idxs[keep]

    return torch.tensor(idxs_out).to(boxes.device), torch.tensor(scores_out).to(scores.device)