File size: 8,714 Bytes
ec9a6bc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from torch import nn
import numpy as np
import kaolin
import tqdm
from pytorch3d.ops.knn import knn_gather, knn_points
from pytorch3d.transforms import so3_exponential_map

from lib.network.MLP import MLP
from lib.network.PositionalEmbedding import get_embedder

from lib.utils.dmtet_utils import marching_tetrahedra

class MeshHeadModule(nn.Module):
    def __init__(self, cfg, init_landmarks_3d_neutral):
        super(MeshHeadModule, self).__init__()
        
        self.geo_mlp = MLP(cfg.geo_mlp, last_op=nn.Tanh())
        self.exp_color_mlp = MLP(cfg.exp_color_mlp, last_op=None)
        self.pose_color_mlp = MLP(cfg.pose_color_mlp, last_op=None)
        self.exp_deform_mlp = MLP(cfg.exp_deform_mlp, last_op=nn.Tanh())
        self.pose_deform_mlp = MLP(cfg.pose_deform_mlp, last_op=nn.Tanh())

        self.landmarks_3d_neutral = nn.Parameter(init_landmarks_3d_neutral)

        self.pos_embedding, _ = get_embedder(cfg.pos_freq)

        self.model_bbox = cfg.model_bbox
        self.dist_threshold_near = cfg.dist_threshold_near
        self.dist_threshold_far = cfg.dist_threshold_far
        self.deform_scale = cfg.deform_scale

        tets_data = np.load('assets/tets_data.npz')
        self.register_buffer('tet_verts', torch.from_numpy(tets_data['tet_verts']))
        self.register_buffer('tets', torch.from_numpy(tets_data['tets']))
        self.grid_res = 128

        if cfg.subdivide:
            self.subdivide()

    def geometry(self, geo_input):
        pred = self.geo_mlp(geo_input)
        return pred

    def exp_color(self, color_input):
        verts_color = self.exp_color_mlp(color_input)
        return verts_color
    
    def pose_color(self, color_input):
        verts_color = self.pose_color_mlp(color_input)
        return verts_color
    
    def exp_deform(self, deform_input):
        deform = self.exp_deform_mlp(deform_input)
        return deform
    
    def pose_deform(self, deform_input):
        deform = self.pose_deform_mlp(deform_input)
        return deform
        
    def get_landmarks(self):
        return self.landmarks_3d_neutral

    def subdivide(self):
        new_tet_verts, new_tets = kaolin.ops.mesh.subdivide_tetmesh(self.tet_verts.unsqueeze(0), self.tets)
        self.tet_verts = new_tet_verts[0]
        self.tets = new_tets
        self.grid_res *= 2

    def reconstruct(self, data):
        B = data['exp_coeff'].shape[0]

        query_pts = self.tet_verts.unsqueeze(0).repeat(B, 1, 1)
        geo_input = self.pos_embedding(query_pts).permute(0, 2, 1)

        pred = self.geometry(geo_input)

        sdf, deform, features = pred[:, :1, :], pred[:, 1:4, :], pred[:, 4:, :]
        sdf = sdf.permute(0, 2, 1)
        features = features.permute(0, 2, 1)
        verts_deformed = (query_pts + torch.tanh(deform.permute(0, 2, 1)) / self.grid_res)
        verts_list, features_list, faces_list = marching_tetrahedra(verts_deformed, features, self.tets, sdf)

        data['verts0_list'] = verts_list
        data['faces_list'] = faces_list

        verts_batch = []
        verts_features_batch = []
        num_pts_max = 0
        for b in range(B):
            if verts_list[b].shape[0] > num_pts_max:
                num_pts_max = verts_list[b].shape[0]
            
        for b in range(B):
            verts_batch.append(torch.cat([verts_list[b], torch.zeros([num_pts_max - verts_list[b].shape[0], verts_list[b].shape[1]], device=verts_list[b].device)], 0))
            verts_features_batch.append(torch.cat([features_list[b], torch.zeros([num_pts_max - features_list[b].shape[0], features_list[b].shape[1]], device=features_list[b].device)], 0))
        verts_batch = torch.stack(verts_batch, 0)
        verts_features_batch = torch.stack(verts_features_batch, 0)

        dists, idx, _ = knn_points(verts_batch, data['landmarks_3d_neutral'])
        exp_weights = torch.clamp((self.dist_threshold_far - dists) / (self.dist_threshold_far - self.dist_threshold_near), 0.0, 1.0)
        pose_weights = 1 - exp_weights

        exp_color_input = torch.cat([verts_features_batch.permute(0, 2, 1), data['exp_coeff'].unsqueeze(-1).repeat(1, 1, num_pts_max)], 1)
        verts_color_batch = self.exp_color(exp_color_input).permute(0, 2, 1) * exp_weights
        
        pose_color_input = torch.cat([verts_features_batch.permute(0, 2, 1), self.pos_embedding(data['pose']).unsqueeze(-1).repeat(1, 1, num_pts_max)], 1)
        verts_color_batch = verts_color_batch + self.pose_color(pose_color_input).permute(0, 2, 1) * pose_weights

        exp_deform_input = torch.cat([self.pos_embedding(verts_batch).permute(0, 2, 1), data['exp_coeff'].unsqueeze(-1).repeat(1, 1, num_pts_max)], 1)
        exp_deform = self.exp_deform(exp_deform_input).permute(0, 2, 1)
        verts_batch = verts_batch + exp_deform * exp_weights * self.deform_scale

        pose_deform_input = torch.cat([self.pos_embedding(verts_batch).permute(0, 2, 1), self.pos_embedding(data['pose']).unsqueeze(-1).repeat(1, 1, num_pts_max)], 1)
        pose_deform = self.pose_deform(pose_deform_input).permute(0, 2, 1)
        verts_batch = verts_batch + pose_deform * pose_weights * self.deform_scale

        if 'pose' in data:
            R = so3_exponential_map(data['pose'][:, :3])
            T = data['pose'][:, None, 3:]
            S = data['scale'][:, :, None]
            verts_batch = torch.bmm(verts_batch * S, R.permute(0, 2, 1)) + T

        data['exp_deform'] = exp_deform
        data['pose_deform'] = pose_deform
        data['verts_list'] = [verts_batch[b, :verts_list[b].shape[0], :] for b in range(B)]
        data['verts_color_list'] = [verts_color_batch[b, :verts_list[b].shape[0], :] for b in range(B)]
        return data
    
    def reconstruct_neutral(self):
        query_pts = self.tet_verts.unsqueeze(0)
        geo_input = self.pos_embedding(query_pts).permute(0, 2, 1)

        pred = self.geometry(geo_input)

        sdf, deform, features = pred[:, :1, :], pred[:, 1:4, :], pred[:, 4:, :]
        sdf = sdf.permute(0, 2, 1)
        features = features.permute(0, 2, 1)
        verts_deformed = (query_pts + torch.tanh(deform.permute(0, 2, 1)) / self.grid_res)
        verts_list, features_list, faces_list = marching_tetrahedra(verts_deformed, features, self.tets, sdf)

        data = {}
        data['verts'] = verts_list[0]
        data['faces'] = faces_list[0]
        data['verts_feature'] = features_list[0]
        return data
    
    def query_sdf(self, data):
        query_pts = data['query_pts']

        geo_input = self.pos_embedding(query_pts).permute(0, 2, 1)

        pred = self.geometry(geo_input)
        sdf = pred[:, :1, :]
        sdf = sdf.permute(0, 2, 1)

        data['sdf'] = sdf
        return data
    
    def deform(self, data):
        exp_coeff = data['exp_coeff']
        query_pts = data['query_pts']
        
        geo_input = self.pos_embedding(query_pts).permute(0, 2, 1)

        pred = self.geometry(geo_input)
        sdf, deform = pred[:, :1, :], pred[:, 1:4, :]
        query_pts = (query_pts + torch.tanh(deform).permute(0, 2, 1) / self.grid_res)

        exp_deform_input = torch.cat([self.pos_embedding(query_pts).permute(0, 2, 1), exp_coeff.unsqueeze(-1).repeat(1, 1, query_pts.shape[1])], 1)
        exp_deform = self.exp_deform(exp_deform_input).permute(0, 2, 1)

        deformed_pts = query_pts + exp_deform * self.deform_scale

        data['deformed_pts'] = deformed_pts
        return data
    
    def in_bbox(self, verts, bbox):
        is_in_bbox = (verts[:, :, 0] > bbox[0][0]) & \
                     (verts[:, :, 1] > bbox[1][0]) & \
                     (verts[:, :, 2] > bbox[2][0]) & \
                     (verts[:, :, 0] < bbox[0][1]) & \
                     (verts[:, :, 1] < bbox[1][1]) & \
                     (verts[:, :, 2] < bbox[2][1])
        return is_in_bbox
    
    def pre_train_sphere(self, iter, device):
        loss_fn = torch.nn.MSELoss()
        optimizer = torch.optim.Adam(list(self.parameters()), lr=1e-3)

        for i in tqdm.tqdm(range(iter)):
            query_pts = torch.rand((8, 1024, 3), device=device) * 3 - 1.5
            ref_value  = torch.sqrt((query_pts**2).sum(-1)) - 1.0
            data = {
                'query_pts': query_pts
                }
            data = self.query_sdf(data)
            sdf = data['sdf']
            loss = loss_fn(sdf[:, :, 0], ref_value)
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

            print("Pre-trained MLP", loss.item())