File size: 7,475 Bytes
bd95c9c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

"""
MHR shape to SMPL shape converter.

This class handles the conversion of MHR shape parameters to SMPL shape parameters.
"""

import argparse
from pathlib import Path

import imageio
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import trimesh
from tqdm import tqdm

from soma.geometry.barycentric_interp import BarycentricInterpolator
from soma.geometry.batched_skinning import BatchedSkinning
from soma.soma import SOMALayer
from tools.vis_pyrender import MeshRenderer, look_at, set_pyopengl_platform


def get_smooth_noise(T, dim, device, num_keyframes=None, mode="normal"):
    if num_keyframes is None:
        num_keyframes = max(3, T // 30)

    if mode == "normal":
        keyframes = torch.randn(1, dim, num_keyframes, device=device)
    elif mode == "uniform":
        keyframes = torch.rand(1, dim, num_keyframes, device=device)

    res = F.interpolate(keyframes, size=T, mode="linear", align_corners=True)[0].T
    return res


class ShapeTransfer(nn.Module):
    """
    MHR shape to SMPL shape converter.
    """

    def __init__(self, data_root, device="cuda"):
        """
        Initialize the shape transfer.
        """
        super().__init__()
        self.data_root = Path(data_root)
        self.device = device

        self.mhr_soma = SOMALayer(
            data_root=data_root, device=device, identity_model_type="mhr", mode="warp"
        )
        self.smpl_soma = SOMALayer(
            data_root=data_root, device=device, identity_model_type="smpl", mode="warp"
        )
        self.soma_to_smpl = self.get_soma_to_smpl_interpolator()

        smpl_rest_mesh = trimesh.load(
            self.data_root / "SMPL" / "smpl_base_body.obj", maintain_order=True, process=False
        )
        smpl_rest_shape = torch.from_numpy(smpl_rest_mesh.vertices).float().to(device).unsqueeze(0)
        self.smpl_rest_shape_soma = self.smpl_soma.identity_model.identity_model_to_soma(
            smpl_rest_shape
        )
        self.posed_world_smpl_tpose = self.mhr_soma.skeleton_transfer.fit(self.smpl_rest_shape_soma)

    def get_soma_to_smpl_interpolator(self):
        mesh_smpl = trimesh.load(
            self.data_root / "SMPL" / "smpl_base_body.obj", maintain_order=True, process=False
        )
        V_smpl = torch.from_numpy(mesh_smpl.vertices).float().to(self.device)
        mesh_nv = trimesh.load(
            self.data_root / "SMPL" / "Nova_wrap.obj", maintain_order=True, process=False
        )
        V_nv = torch.from_numpy(mesh_nv.vertices).float().to(self.device)
        F_nv = torch.from_numpy(mesh_nv.faces).int().to(self.device)
        return BarycentricInterpolator(V_nv, F_nv, V_smpl)

    def forward(self, identity_coeffs, scale_params):
        """
        Forward pass.
        """
        batch_size = identity_coeffs.shape[0]
        device = identity_coeffs.device

        # 1. Get MHR rest shape
        mhr_rest_shape_soma = self.mhr_soma.identity_model(identity_coeffs, scale_params)

        # 2. Get MHR posed world transforms
        posed_world_mhr = self.mhr_soma.skeleton_transfer.fit(mhr_rest_shape_soma)

        # 3. Skin the MHR rest shape to get it under the SMPL rest pose
        batched_skinning = BatchedSkinning(
            self.mhr_soma.joint_parent_ids,
            self.mhr_soma.skinning_weights,
            posed_world_mhr,
            mhr_rest_shape_soma,
            joint_orient=self.posed_world_smpl_tpose[0],
            mode=self.mhr_soma.mode,
        )

        pose_rotations = torch.eye(3, device=device).unsqueeze(0).expand(batch_size, 78, 3, 3)
        pose_translations = self.posed_world_smpl_tpose[:, 1, :3, 3].expand(batch_size, 3)

        vertices, T_world = batched_skinning.pose(
            local_rotations=pose_rotations,
            hips_translations=pose_translations,
            return_transforms=True,
        )

        # 4. Get the SMPL topology vertices
        mhr_vertices_smpl = self.soma_to_smpl(vertices)

        # 5. Solve the betas
        B = (
            mhr_vertices_smpl - self.smpl_soma.identity_model.identity_model.v_template[None]
        ).reshape(batch_size, -1)
        A = self.smpl_soma.identity_model.identity_model.shape_dirs.reshape(-1, 10)

        betas = torch.linalg.lstsq(A, B.T).solution.T

        return betas


# Example usage
if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Shape transfer.")
    parser.add_argument("--data_root", type=str, default="./assets", help="Path to the data root.")
    parser.add_argument("--device", default="cuda:0")
    parser.add_argument("--output-dir", default="out/")
    parser.add_argument("--image-size", type=int, default=1920)
    parser.add_argument("--sequence-length", type=int, default=300)
    parser.add_argument("--pyopengl-platform", default="osmesa")
    args = parser.parse_args()

    set_pyopengl_platform(args.pyopengl_platform)

    device = "cuda"
    shape_transfer = ShapeTransfer(args.data_root, device)
    T = args.sequence_length

    mhr_im = shape_transfer.mhr_soma.identity_model
    identity_coeffs = get_smooth_noise(T, mhr_im.num_identity_coeffs, device)
    scale_params = get_smooth_noise(T, mhr_im.num_scale_params, device, mode="normal") * 0.2
    zero_pose = torch.zeros(1, 77, 3, device=device)
    zero_transl = torch.zeros(1, 3, device=device)

    betas = shape_transfer(identity_coeffs, scale_params)

    smpl_vertices = shape_transfer.smpl_soma(zero_pose, betas, None, zero_transl)["vertices"]
    mhr_vertices = shape_transfer.mhr_soma(zero_pose, identity_coeffs, scale_params, zero_transl)[
        "vertices"
    ]

    smpl_vertices = smpl_vertices.detach().cpu().numpy()
    mhr_vertices = mhr_vertices.detach().cpu().numpy()
    faces = shape_transfer.mhr_soma.faces.cpu().numpy()

    print("Rendering videos...")
    colors = {
        "mhr": (0.98, 0.65, 0.15, 1.0),
        "anny": (0.25, 0.75, 1.0, 1.0),
        "smpl": (0.55, 0.15, 0.85, 1.0),
    }

    def save_video(frames, path, fps=30):
        imageio.mimsave(path, frames, fps=fps)
        print(f"Saved {path}")

    renderer = MeshRenderer(image_size=args.image_size, light_intensity=5)

    cam_pose = look_at(
        eye=np.array([0.0, 0.0, 6.0]),
        target=np.array([0.0, 0.0, 0.0]),
        up=np.array([0.0, 1.0, 0.0]),
    )
    light_dir = np.array([0.0, -0.5, -1.0])

    faces = shape_transfer.mhr_soma.faces.cpu().numpy()
    frames = []

    for t in tqdm(range(T)):
        mhr_img = renderer.render(
            mhr_vertices[t],
            faces,
            mesh_color=colors["mhr"],
            cam_pose=cam_pose,
            light_dir=light_dir,
            metallic=0.0,
            roughness=0.5,
            base_color_factor=[0.9, 0.9, 0.9, 1.0],
        )
        smpl_img = renderer.render(
            smpl_vertices[t],
            faces,
            mesh_color=colors["smpl"],
            cam_pose=cam_pose,
            light_dir=light_dir,
            metallic=0.0,
            roughness=0.5,
            base_color_factor=[0.9, 0.9, 0.9, 1.0],
        )
        merged_img = (0.5 * mhr_img + 0.5 * smpl_img).astype(np.uint8)
        img = np.concatenate([mhr_img, merged_img, smpl_img], axis=1)
        frames.append(img[..., ::-1])

    renderer.delete()
    save_video(frames, Path(args.output_dir) / "shape_transfer.mp4")