Skip to content

LMDS

HerdNetLMDS

Bases: LMDS

Source code in PytorchWildlife/models/detection/herdnet/animaloc/eval/lmds.py 
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
class HerdNetLMDS(LMDS):

    def __init__(
        self, 
        up: bool = True, 
        kernel_size: tuple = (3,3), 
        adapt_ts: float = 0.3, 
        neg_ts: float = 0.1
        ) -> None:
        '''
        Args:
            up (bool, optional): set to False to disable class maps upsampling.
                Defaults to True.
            kernel_size (tuple, optional): size of the kernel used to select local
                maxima. Defaults to (3,3) (as in the paper).
            adapt_ts (float, optional): adaptive threshold to select final points
                from candidates. Defaults to 0.3.
            neg_ts (float, optional): negative sample threshold used to define if 
                an image is a negative sample or not. Defaults to 0.1 (as in the paper).
        '''

        super().__init__(kernel_size=kernel_size, adapt_ts=adapt_ts, neg_ts=neg_ts)

        self.up = up

    def __call__(self, outputs: List[torch.Tensor]) -> Tuple[list, list, list, list, list]:
        """
        Args:
            outputs (List[torch.Tensor]): Outputs of HerdNet, i.e., 2 tensors:
                - heatmap: [B,1,H,W],
                - class map: [B,C,H/16,W/16].

        Returns:
            Tuple[list, list, list, list, list]:
                Counts, locations, labels, class scores, and detection scores per batch.
        """

        heatmap, clsmap = outputs

        # upsample class map
        if self.up:
            scale_factor = 16
            clsmap = F.interpolate(clsmap, scale_factor=scale_factor, mode='nearest')

        # softmax
        cls_scores = torch.softmax(clsmap, dim=1)[:,1:,:,:]

        # cat to heatmap
        outmaps = torch.cat([heatmap, cls_scores], dim=1)

        # LMDS
        batch_size, channels = outmaps.shape[:2]

        b_counts, b_labels, b_scores, b_locs, b_dscores = [], [], [], [], []
        for b in range(batch_size):

            _, locs, _ = self._lmds(heatmap[b][0])

            cls_idx = torch.argmax(clsmap[b,1:,:,:], dim=0)
            classes = torch.add(cls_idx, 1)

            h_idx = torch.Tensor([l[0] for l in locs]).long()
            w_idx = torch.Tensor([l[1] for l in locs]).long()
            labels = classes[h_idx, w_idx].long().tolist()

            chan_idx = cls_idx[h_idx, w_idx].long().tolist()
            scores = cls_scores[b, chan_idx, h_idx, w_idx].float().tolist()

            dscores = heatmap[b, 0, h_idx, w_idx].float().tolist()

            counts = [labels.count(i) for i in range(1, channels)]

            b_labels.append(labels)
            b_scores.append(scores)
            b_locs.append(locs)
            b_counts.append(counts)
            b_dscores.append(dscores)

        return b_counts, b_locs, b_labels, b_scores, b_dscores

__call__(outputs)

Parameters:

Name Type Description Default
outputs List[Tensor]

Outputs of HerdNet, i.e., 2 tensors: - heatmap: [B,1,H,W], - class map: [B,C,H/16,W/16].

 required

Returns:

Type Description
Tuple[list, list, list, list, list]

Tuple[list, list, list, list, list]: Counts, locations, labels, class scores, and detection scores per batch.
Source code in PytorchWildlife/models/detection/herdnet/animaloc/eval/lmds.py 
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
def __call__(self, outputs: List[torch.Tensor]) -> Tuple[list, list, list, list, list]:
    """
    Args:
        outputs (List[torch.Tensor]): Outputs of HerdNet, i.e., 2 tensors:
            - heatmap: [B,1,H,W],
            - class map: [B,C,H/16,W/16].

    Returns:
        Tuple[list, list, list, list, list]:
            Counts, locations, labels, class scores, and detection scores per batch.
    """

    heatmap, clsmap = outputs

    # upsample class map
    if self.up:
        scale_factor = 16
        clsmap = F.interpolate(clsmap, scale_factor=scale_factor, mode='nearest')

    # softmax
    cls_scores = torch.softmax(clsmap, dim=1)[:,1:,:,:]

    # cat to heatmap
    outmaps = torch.cat([heatmap, cls_scores], dim=1)

    # LMDS
    batch_size, channels = outmaps.shape[:2]

    b_counts, b_labels, b_scores, b_locs, b_dscores = [], [], [], [], []
    for b in range(batch_size):

        _, locs, _ = self._lmds(heatmap[b][0])

        cls_idx = torch.argmax(clsmap[b,1:,:,:], dim=0)
        classes = torch.add(cls_idx, 1)

        h_idx = torch.Tensor([l[0] for l in locs]).long()
        w_idx = torch.Tensor([l[1] for l in locs]).long()
        labels = classes[h_idx, w_idx].long().tolist()

        chan_idx = cls_idx[h_idx, w_idx].long().tolist()
        scores = cls_scores[b, chan_idx, h_idx, w_idx].float().tolist()

        dscores = heatmap[b, 0, h_idx, w_idx].float().tolist()

        counts = [labels.count(i) for i in range(1, channels)]

        b_labels.append(labels)
        b_scores.append(scores)
        b_locs.append(locs)
        b_counts.append(counts)
        b_dscores.append(dscores)

    return b_counts, b_locs, b_labels, b_scores, b_dscores

__init__(up=True, kernel_size=(3, 3), adapt_ts=0.3, neg_ts=0.1)

Parameters:

Name Type Description Default
up bool

set to False to disable class maps upsampling. Defaults to True.

 True
kernel_size tuple

size of the kernel used to select local maxima. Defaults to (3,3) (as in the paper).

 (3, 3)
adapt_ts float

adaptive threshold to select final points from candidates. Defaults to 0.3.

 0.3
neg_ts float

negative sample threshold used to define if an image is a negative sample or not. Defaults to 0.1 (as in the paper).

 0.1
Source code in PytorchWildlife/models/detection/herdnet/animaloc/eval/lmds.py 
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
def __init__(
    self, 
    up: bool = True, 
    kernel_size: tuple = (3,3), 
    adapt_ts: float = 0.3, 
    neg_ts: float = 0.1
    ) -> None:
    '''
    Args:
        up (bool, optional): set to False to disable class maps upsampling.
            Defaults to True.
        kernel_size (tuple, optional): size of the kernel used to select local
            maxima. Defaults to (3,3) (as in the paper).
        adapt_ts (float, optional): adaptive threshold to select final points
            from candidates. Defaults to 0.3.
        neg_ts (float, optional): negative sample threshold used to define if 
            an image is a negative sample or not. Defaults to 0.1 (as in the paper).
    '''

    super().__init__(kernel_size=kernel_size, adapt_ts=adapt_ts, neg_ts=neg_ts)

    self.up = up

LMDS

Local Maxima Detection Strategy

Adapted and enhanced from https://github.com/dk-liang/FIDTM (author: dklinag) available under the MIT license

Source code in PytorchWildlife/models/detection/herdnet/animaloc/eval/lmds.py 
 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
class LMDS:
    ''' Local Maxima Detection Strategy 

    Adapted and enhanced from https://github.com/dk-liang/FIDTM (author: dklinag)
    available under the MIT license '''

    def __init__(
        self, 
        kernel_size: tuple = (3,3),
        adapt_ts: float = 100.0/255.0, 
        neg_ts: float = 0.1
        ) -> None:
        '''
        Args:
            kernel_size (tuple, optional): size of the kernel used to select local
                maxima. Defaults to (3,3) (as in the paper).
            adapt_ts (float, optional): adaptive threshold to select final points
                from candidates. Defaults to 100.0/255.0 (as in the paper).
            neg_ts (float, optional): negative sample threshold used to define if 
                an image is a negative sample or not. Defaults to 0.1 (as in the paper).
        '''

        assert kernel_size[0] == kernel_size[1], \
            f'The kernel shape must be a square, got {kernel_size[0]}x{kernel_size[1]}'
        assert not kernel_size[0] % 2 == 0, \
            f'The kernel size must be odd, got {kernel_size[0]}'

        self.kernel_size = tuple(kernel_size)
        self.adapt_ts = adapt_ts
        self.neg_ts = neg_ts

    def __call__(self, est_map: torch.Tensor) -> Tuple[list,list,list,list]:
        '''
        Args:
            est_map (torch.Tensor): the estimated FIDT map

        Returns:
            Tuple[list,list,list,list]
                counts, labels, scores and locations per batch
        '''
        batch_size, classes = est_map.shape[:2]

        b_counts, b_labels, b_scores, b_locs = [], [], [], []
        for b in range(batch_size):
            counts, labels, scores, locs = [], [], [], []

            for c in range(classes):
                count, loc, score = self._lmds(est_map[b][c])
                counts.append(count)
                labels = [*labels, *[c+1]*count]
                scores = [*scores, *score]
                locs = [*locs, *loc]

            b_counts.append(counts)
            b_labels.append(labels)
            b_scores.append(scores)
            b_locs.append(locs)

        return b_counts, b_locs, b_labels, b_scores

    def _local_max(self, est_map: torch.Tensor) -> torch.Tensor:
        ''' Shape: est_map = [B,C,H,W] '''

        pad = int(self.kernel_size[0] / 2)
        keep = torch.nn.functional.max_pool2d(est_map, kernel_size=self.kernel_size, stride=1, padding=pad)
        keep = (keep == est_map).float()
        est_map = keep * est_map

        return est_map

    def _get_locs_and_scores(
        self, 
        locs_map: torch.Tensor, 
        scores_map: torch.Tensor
        ) -> Tuple[torch.Tensor, torch.Tensor]:
        ''' Shapes: locs_map = [H,W] and scores_map = [H,W] '''

        locs_map = locs_map.data.cpu().numpy()
        scores_map = scores_map.data.cpu().numpy()
        locs = []
        scores = []
        for i, j in numpy.argwhere(locs_map ==  1):
            locs.append((i,j))
            scores.append(scores_map[i][j])

        return torch.Tensor(locs), torch.Tensor(scores)

    def _lmds(self, est_map: torch.Tensor) -> Tuple[int, list, list]:
        ''' Shape: est_map = [H,W] '''

        est_map_max = torch.max(est_map).item()

        # local maxima
        est_map = self._local_max(est_map.unsqueeze(0).unsqueeze(0))

        # adaptive threshold for counting
        est_map[est_map < self.adapt_ts * est_map_max] = 0
        scores_map = torch.clone(est_map)
        est_map[est_map > 0] = 1

        # negative sample
        if est_map_max < self.neg_ts:
            est_map = est_map * 0

        # count
        count = int(torch.sum(est_map).item())

        # locations and scores
        locs, scores = self._get_locs_and_scores(
            est_map.squeeze(0).squeeze(0), 
            scores_map.squeeze(0).squeeze(0)
            )

        return count, locs.tolist(), scores.tolist()

__call__(est_map)

Parameters:

Name Type Description Default
est_map Tensor

the estimated FIDT map

 required

Returns:

Type Description
Tuple[list, list, list, list]

Tuple[list,list,list,list] counts, labels, scores and locations per batch
Source code in PytorchWildlife/models/detection/herdnet/animaloc/eval/lmds.py 
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
def __call__(self, est_map: torch.Tensor) -> Tuple[list,list,list,list]:
    '''
    Args:
        est_map (torch.Tensor): the estimated FIDT map

    Returns:
        Tuple[list,list,list,list]
            counts, labels, scores and locations per batch
    '''
    batch_size, classes = est_map.shape[:2]

    b_counts, b_labels, b_scores, b_locs = [], [], [], []
    for b in range(batch_size):
        counts, labels, scores, locs = [], [], [], []

        for c in range(classes):
            count, loc, score = self._lmds(est_map[b][c])
            counts.append(count)
            labels = [*labels, *[c+1]*count]
            scores = [*scores, *score]
            locs = [*locs, *loc]

        b_counts.append(counts)
        b_labels.append(labels)
        b_scores.append(scores)
        b_locs.append(locs)

    return b_counts, b_locs, b_labels, b_scores

__init__(kernel_size=(3, 3), adapt_ts=100.0 / 255.0, neg_ts=0.1)

Parameters:

Name Type Description Default
kernel_size tuple

size of the kernel used to select local maxima. Defaults to (3,3) (as in the paper).

 (3, 3)
adapt_ts float

adaptive threshold to select final points from candidates. Defaults to 100.0/255.0 (as in the paper).

 100.0 / 255.0
neg_ts float

negative sample threshold used to define if an image is a negative sample or not. Defaults to 0.1 (as in the paper).

 0.1
Source code in PytorchWildlife/models/detection/herdnet/animaloc/eval/lmds.py 
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
def __init__(
    self, 
    kernel_size: tuple = (3,3),
    adapt_ts: float = 100.0/255.0, 
    neg_ts: float = 0.1
    ) -> None:
    '''
    Args:
        kernel_size (tuple, optional): size of the kernel used to select local
            maxima. Defaults to (3,3) (as in the paper).
        adapt_ts (float, optional): adaptive threshold to select final points
            from candidates. Defaults to 100.0/255.0 (as in the paper).
        neg_ts (float, optional): negative sample threshold used to define if 
            an image is a negative sample or not. Defaults to 0.1 (as in the paper).
    '''

    assert kernel_size[0] == kernel_size[1], \
        f'The kernel shape must be a square, got {kernel_size[0]}x{kernel_size[1]}'
    assert not kernel_size[0] % 2 == 0, \
        f'The kernel size must be odd, got {kernel_size[0]}'

    self.kernel_size = tuple(kernel_size)
    self.adapt_ts = adapt_ts
    self.neg_ts = neg_ts