File size: 13,426 Bytes
404d2af
 
 
8b973ee
404d2af
e430362
8b973ee
8f7d727
8b973ee
404d2af
 
 
 
 
 
 
 
 
8b973ee
 
404d2af
 
e430362
 
 
 
8b973ee
404d2af
 
8b973ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
404d2af
8b973ee
e430362
8b973ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
404d2af
8b973ee
 
 
 
404d2af
 
8b973ee
 
 
 
 
 
 
 
404d2af
 
 
8b973ee
404d2af
 
 
 
 
 
 
 
 
 
8b973ee
404d2af
 
 
 
 
 
 
8b973ee
404d2af
8b973ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
404d2af
8b973ee
 
 
 
 
 
 
 
 
404d2af
 
 
 
8b973ee
404d2af
8b973ee
 
404d2af
8b973ee
 
 
 
 
 
 
 
 
 
404d2af
8b973ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
404d2af
 
 
8b973ee
404d2af
8b973ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
404d2af
8b973ee
404d2af
8b973ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
404d2af
8b973ee
 
 
 
 
404d2af
8b973ee
 
 
 
 
 
 
 
404d2af
 
8b973ee
404d2af
8b973ee
 
 
 
 
 
 
 
 
 
404d2af
8b973ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
404d2af
8b973ee
 
 
 
404d2af
8b973ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
import torch
import torch.nn as nn

eps = 1e-8

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


def sinkhorn(M, r, c, iteration):
    p = torch.softmax(M, dim=-1)
    u = torch.ones_like(r)
    v = torch.ones_like(c)
    for _ in range(iteration):
        u = r / ((p * v.unsqueeze(-2)).sum(-1) + eps)
        v = c / ((p * u.unsqueeze(-1)).sum(-2) + eps)
    p = p * u.unsqueeze(-1) * v.unsqueeze(-2)
    return p


def sink_algorithm(M, dustbin, iteration):
    M = torch.cat([M, dustbin.expand([M.shape[0], M.shape[1], 1])], dim=-1)
    M = torch.cat([M, dustbin.expand([M.shape[0], 1, M.shape[2]])], dim=-2)
    r = torch.ones([M.shape[0], M.shape[1] - 1], device=device)
    r = torch.cat([r, torch.ones([M.shape[0], 1], device=device) * M.shape[1]], dim=-1)
    c = torch.ones([M.shape[0], M.shape[2] - 1], device=device)
    c = torch.cat([c, torch.ones([M.shape[0], 1], device=device) * M.shape[2]], dim=-1)
    p = sinkhorn(M, r, c, iteration)
    return p


def seeding(
    nn_index1,
    nn_index2,
    x1,
    x2,
    topk,
    match_score,
    confbar,
    nms_radius,
    use_mc=True,
    test=False,
):

    # apply mutual check before nms
    if use_mc:
        mask_not_mutual = nn_index2.gather(dim=-1, index=nn_index1) != torch.arange(
            nn_index1.shape[1], device=device
        )
        match_score[mask_not_mutual] = -1
    # NMS
    pos_dismat1 = (
        (
            (x1.norm(p=2, dim=-1) ** 2).unsqueeze_(-1)
            + (x1.norm(p=2, dim=-1) ** 2).unsqueeze_(-2)
            - 2 * (x1 @ x1.transpose(1, 2))
        )
        .abs_()
        .sqrt_()
    )
    x2 = x2.gather(index=nn_index1.unsqueeze(-1).expand(-1, -1, 2), dim=1)
    pos_dismat2 = (
        (
            (x2.norm(p=2, dim=-1) ** 2).unsqueeze_(-1)
            + (x2.norm(p=2, dim=-1) ** 2).unsqueeze_(-2)
            - 2 * (x2 @ x2.transpose(1, 2))
        )
        .abs_()
        .sqrt_()
    )
    radius1, radius2 = nms_radius * pos_dismat1.mean(
        dim=(1, 2), keepdim=True
    ), nms_radius * pos_dismat2.mean(dim=(1, 2), keepdim=True)
    nms_mask = (pos_dismat1 >= radius1) & (pos_dismat2 >= radius2)
    mask_not_local_max = (
        match_score.unsqueeze(-1) >= match_score.unsqueeze(-2)
    ) | nms_mask
    mask_not_local_max = ~(mask_not_local_max.min(dim=-1).values)
    match_score[mask_not_local_max] = -1

    # confidence bar
    match_score[match_score < confbar] = -1
    mask_survive = match_score > 0
    if test:
        topk = min(mask_survive.sum(dim=1)[0] + 2, topk)
    _, topindex = torch.topk(match_score, topk, dim=-1)  # b*k
    seed_index1, seed_index2 = topindex, nn_index1.gather(index=topindex, dim=-1)
    return seed_index1, seed_index2


class PointCN(nn.Module):
    def __init__(self, channels, out_channels):
        nn.Module.__init__(self)
        self.shot_cut = nn.Conv1d(channels, out_channels, kernel_size=1)
        self.conv = nn.Sequential(
            nn.InstanceNorm1d(channels, eps=1e-3),
            nn.SyncBatchNorm(channels),
            nn.ReLU(),
            nn.Conv1d(channels, channels, kernel_size=1),
            nn.InstanceNorm1d(channels, eps=1e-3),
            nn.SyncBatchNorm(channels),
            nn.ReLU(),
            nn.Conv1d(channels, out_channels, kernel_size=1),
        )

    def forward(self, x):
        return self.conv(x) + self.shot_cut(x)


class attention_propagantion(nn.Module):
    def __init__(self, channel, head):
        nn.Module.__init__(self)
        self.head = head
        self.head_dim = channel // head
        self.query_filter, self.key_filter, self.value_filter = (
            nn.Conv1d(channel, channel, kernel_size=1),
            nn.Conv1d(channel, channel, kernel_size=1),
            nn.Conv1d(channel, channel, kernel_size=1),
        )
        self.mh_filter = nn.Conv1d(channel, channel, kernel_size=1)
        self.cat_filter = nn.Sequential(
            nn.Conv1d(2 * channel, 2 * channel, kernel_size=1),
            nn.SyncBatchNorm(2 * channel),
            nn.ReLU(),
            nn.Conv1d(2 * channel, channel, kernel_size=1),
        )

    def forward(self, desc1, desc2, weight_v=None):
        # desc1(q) attend to desc2(k,v)
        batch_size = desc1.shape[0]
        query, key, value = (
            self.query_filter(desc1).view(batch_size, self.head, self.head_dim, -1),
            self.key_filter(desc2).view(batch_size, self.head, self.head_dim, -1),
            self.value_filter(desc2).view(batch_size, self.head, self.head_dim, -1),
        )
        if weight_v is not None:
            value = value * weight_v.view(batch_size, 1, 1, -1)
        score = torch.softmax(
            torch.einsum("bhdn,bhdm->bhnm", query, key) / self.head_dim**0.5, dim=-1
        )
        add_value = torch.einsum("bhnm,bhdm->bhdn", score, value).reshape(
            batch_size, self.head_dim * self.head, -1
        )
        add_value = self.mh_filter(add_value)
        desc1_new = desc1 + self.cat_filter(torch.cat([desc1, add_value], dim=1))
        return desc1_new


class hybrid_block(nn.Module):
    def __init__(self, channel, head):
        nn.Module.__init__(self)
        self.head = head
        self.channel = channel
        self.attention_block_down = attention_propagantion(channel, head)
        self.cluster_filter = nn.Sequential(
            nn.Conv1d(2 * channel, 2 * channel, kernel_size=1),
            nn.SyncBatchNorm(2 * channel),
            nn.ReLU(),
            nn.Conv1d(2 * channel, 2 * channel, kernel_size=1),
        )
        self.cross_filter = attention_propagantion(channel, head)
        self.confidence_filter = PointCN(2 * channel, 1)
        self.attention_block_self = attention_propagantion(channel, head)
        self.attention_block_up = attention_propagantion(channel, head)

    def forward(self, desc1, desc2, seed_index1, seed_index2):
        cluster1, cluster2 = desc1.gather(
            dim=-1, index=seed_index1.unsqueeze(1).expand(-1, self.channel, -1)
        ), desc2.gather(
            dim=-1, index=seed_index2.unsqueeze(1).expand(-1, self.channel, -1)
        )

        # pooling
        cluster1, cluster2 = self.attention_block_down(
            cluster1, desc1
        ), self.attention_block_down(cluster2, desc2)
        concate_cluster = self.cluster_filter(torch.cat([cluster1, cluster2], dim=1))
        # filtering
        cluster1, cluster2 = self.cross_filter(
            concate_cluster[:, : self.channel], concate_cluster[:, self.channel :]
        ), self.cross_filter(
            concate_cluster[:, self.channel :], concate_cluster[:, : self.channel]
        )
        cluster1, cluster2 = self.attention_block_self(
            cluster1, cluster1
        ), self.attention_block_self(cluster2, cluster2)
        # unpooling
        seed_weight = self.confidence_filter(torch.cat([cluster1, cluster2], dim=1))
        seed_weight = torch.sigmoid(seed_weight).squeeze(1)
        desc1_new, desc2_new = self.attention_block_up(
            desc1, cluster1, seed_weight
        ), self.attention_block_up(desc2, cluster2, seed_weight)
        return desc1_new, desc2_new, seed_weight


class matcher(nn.Module):
    def __init__(self, config):
        nn.Module.__init__(self)
        self.seed_top_k = config.seed_top_k
        self.conf_bar = config.conf_bar
        self.seed_radius_coe = config.seed_radius_coe
        self.use_score_encoding = config.use_score_encoding
        self.detach_iter = config.detach_iter
        self.seedlayer = config.seedlayer
        self.layer_num = config.layer_num
        self.sink_iter = config.sink_iter

        self.position_encoder = nn.Sequential(
            nn.Conv1d(3, 32, kernel_size=1)
            if config.use_score_encoding
            else nn.Conv1d(2, 32, kernel_size=1),
            nn.SyncBatchNorm(32),
            nn.ReLU(),
            nn.Conv1d(32, 64, kernel_size=1),
            nn.SyncBatchNorm(64),
            nn.ReLU(),
            nn.Conv1d(64, 128, kernel_size=1),
            nn.SyncBatchNorm(128),
            nn.ReLU(),
            nn.Conv1d(128, 256, kernel_size=1),
            nn.SyncBatchNorm(256),
            nn.ReLU(),
            nn.Conv1d(256, config.net_channels, kernel_size=1),
        )

        self.hybrid_block = nn.Sequential(
            *[
                hybrid_block(config.net_channels, config.head)
                for _ in range(config.layer_num)
            ]
        )
        self.final_project = nn.Conv1d(
            config.net_channels, config.net_channels, kernel_size=1
        )
        self.dustbin = nn.Parameter(torch.tensor(1.5, dtype=torch.float32))

        # if reseeding
        if len(config.seedlayer) != 1:
            self.mid_dustbin = nn.ParameterDict(
                {
                    str(i): nn.Parameter(torch.tensor(2, dtype=torch.float32))
                    for i in config.seedlayer[1:]
                }
            )
            self.mid_final_project = nn.Conv1d(
                config.net_channels, config.net_channels, kernel_size=1
            )

    def forward(self, data, test_mode=True):
        x1, x2, desc1, desc2 = (
            data["x1"][:, :, :2],
            data["x2"][:, :, :2],
            data["desc1"],
            data["desc2"],
        )
        desc1, desc2 = torch.nn.functional.normalize(
            desc1, dim=-1
        ), torch.nn.functional.normalize(desc2, dim=-1)
        if test_mode:
            encode_x1, encode_x2 = data["x1"], data["x2"]
        else:
            encode_x1, encode_x2 = data["aug_x1"], data["aug_x2"]

        # preparation
        desc_dismat = (2 - 2 * torch.matmul(desc1, desc2.transpose(1, 2))).sqrt_()
        values, nn_index = torch.topk(
            desc_dismat, k=2, largest=False, dim=-1, sorted=True
        )
        nn_index2 = torch.min(desc_dismat, dim=1).indices.squeeze(1)
        inverse_ratio_score, nn_index1 = (
            values[:, :, 1] / values[:, :, 0],
            nn_index[:, :, 0],
        )  # get inverse score

        # initial seeding
        seed_index1, seed_index2 = seeding(
            nn_index1,
            nn_index2,
            x1,
            x2,
            self.seed_top_k[0],
            inverse_ratio_score,
            self.conf_bar[0],
            self.seed_radius_coe,
            test=test_mode,
        )

        # position encoding
        desc1, desc2 = desc1.transpose(1, 2), desc2.transpose(1, 2)
        if not self.use_score_encoding:
            encode_x1, encode_x2 = encode_x1[:, :, :2], encode_x2[:, :, :2]
        encode_x1, encode_x2 = encode_x1.transpose(1, 2), encode_x2.transpose(1, 2)
        x1_pos_embedding, x2_pos_embedding = self.position_encoder(
            encode_x1
        ), self.position_encoder(encode_x2)
        aug_desc1, aug_desc2 = x1_pos_embedding + desc1, x2_pos_embedding + desc2

        seed_weight_tower, mid_p_tower, seed_index_tower, nn_index_tower = (
            [],
            [],
            [],
            [],
        )
        seed_index_tower.append(torch.stack([seed_index1, seed_index2], dim=-1))
        nn_index_tower.append(nn_index1)

        seed_para_index = 0
        for i in range(self.layer_num):
            # mid seeding
            if i in self.seedlayer and i != 0:
                seed_para_index += 1
                aug_desc1, aug_desc2 = self.mid_final_project(
                    aug_desc1
                ), self.mid_final_project(aug_desc2)
                M = torch.matmul(aug_desc1.transpose(1, 2), aug_desc2)
                p = sink_algorithm(
                    M, self.mid_dustbin[str(i)], self.sink_iter[seed_para_index - 1]
                )
                mid_p_tower.append(p)
                # rematching with p
                values, nn_index = torch.topk(p[:, :-1, :-1], k=1, dim=-1)
                nn_index2 = torch.max(p[:, :-1, :-1], dim=1).indices.squeeze(1)
                p_match_score, nn_index1 = values[:, :, 0], nn_index[:, :, 0]
                # reseeding
                seed_index1, seed_index2 = seeding(
                    nn_index1,
                    nn_index2,
                    x1,
                    x2,
                    self.seed_top_k[seed_para_index],
                    p_match_score,
                    self.conf_bar[seed_para_index],
                    self.seed_radius_coe,
                    test=test_mode,
                )
                seed_index_tower.append(
                    torch.stack([seed_index1, seed_index2], dim=-1)
                ), nn_index_tower.append(nn_index1)
                if not test_mode and data["step"] < self.detach_iter:
                    aug_desc1, aug_desc2 = aug_desc1.detach(), aug_desc2.detach()

            aug_desc1, aug_desc2, seed_weight = self.hybrid_block[i](
                aug_desc1, aug_desc2, seed_index1, seed_index2
            )
            seed_weight_tower.append(seed_weight)

        aug_desc1, aug_desc2 = self.final_project(aug_desc1), self.final_project(
            aug_desc2
        )
        cmat = torch.matmul(aug_desc1.transpose(1, 2), aug_desc2)
        p = sink_algorithm(cmat, self.dustbin, self.sink_iter[-1])
        # seed_weight_tower: l*b*k
        # seed_index_tower: l*b*k*2
        # nn_index_tower: seed_l*b
        return {
            "p": p,
            "seed_conf": seed_weight_tower,
            "seed_index": seed_index_tower,
            "mid_p": mid_p_tower,
            "nn_index": nn_index_tower,
        }