从UV位置图获得3D人脸

忘是亡心i 2022-10-18 11:30 67阅读 0赞

1. 这部分主要写从UV位置图获得3D人脸，并不包括如何产生UV位置图。

我已经获取一张图片的UV位置图，如下图：

![20200323133017663.jpg][]![20200323133042951.jpg][]![20200323133104898.jpg][]

（a. 裁剪后的人脸区域）                             （b. UV位置图 （仅用来展示的））            （c. UV纹理图）

a, b, c 三张图片shape都是(255, 255, 3) ， UV位置图 b 并不是真正的 UV位置图，  只是将UV位置图转化成图像显示出来。

接下来， 利用a, b 获取 a中人脸对应的3D模型。

2. 主要代码

"""
    如何 从 Position Map 产生 3D Face
    """
    import numpy as np
    import os
    from skimage.io import imread, imsave
     
    # 类似 超参数
    uv_kpt_ind = np.loadtxt("datas/uv-data/uv_kpt_ind.txt").astype(np.int32)  # 2 x 68 get kpt
    face_ind = np.loadtxt("datas/uv-data/face_ind.txt").astype(np.int32)  # get valid vertices in the pos map
    triangles = np.loadtxt("datas/uv-data/triangles.txt").astype(np.int32)  # ntri x 3
     
     
    # 测试一个例子
    face_url = "datas/image00050.jpg"
    face_texture_url = "datas/image00050_tex.jpg"
    # Label 是 npy 数据， 可不是  position map图像(只是用来显示看的)
    face_posmap_url = "datas/image00050.npy"  #  真正的 UV Position Map
     
    global resolution_op
    resolution_op = 256
    # 设置参数
    uv_h = uv_w = 256
    image_h = image_w = 256
     
    image_posmap = np.load(face_posmap_url)
    print(image_posmap.shape)
    print(np.max(image_posmap))
     
     
    # 从 Position Map 获取 顶点
    def get_vertices(pos):
        '''
        Args:
            pos: the 3D position map. shape = (256, 256, 3).
        Returns:
            vertices: the vertices(point cloud). shape = (num of points, 3). n is about 40K here.
        '''
        all_vertices = np.reshape(pos, [resolution_op ** 2, -1])
        vertices = all_vertices[face_ind, :]  # face_ind 是什么呢？
     
        return vertices
     
     
    def get_landmarks(pos):
        """
        Args:
            pos: the 3D position map. shape = (256, 256, 3).
        Returns:
            kpt: 68 3D landmarks. shape = (68, 3).
        :param pos:
        :return:
        """
        kpt = pos[uv_kpt_ind[1, :], uv_kpt_ind[0, :], :]
        return kpt
     
     
    kpt = get_landmarks(image_posmap)
    print(kpt.shape)  # (68, 3)  68个关键点
     
    vertices = get_vertices(image_posmap)
    print(vertices.shape)  # (43867, 3) PRNet 人脸是 43867 个顶点
     
     
    # 保存顶点  不保存纹理
    def dump_to_ply(vertex, tri, wfp):
        header = """ply
        format ascii 1.0
        element vertex {}
        property float x
        property float y
        property float z
        element face {}
        property list uchar int vertex_indices
        end_header"""
        n_vertex = vertex.shape[1]  # ((3, 43867))
        n_face = tri.shape[1]   # ((3, 86906))
        header = header.format(n_vertex, n_face)
     
        with open(wfp, 'w') as f:
            f.write(header + '\n')
            for i in range(n_vertex):  # 顶点
                x, y, z = vertex[:, i]
                f.write('{:.4f} {:.4f} {:.4f}\n'.format(x, y, z))
            for i in range(n_face):  # 三角形
                idx1, idx2, idx3 = tri[:, i]
                f.write('3 {} {} {}\n'.format(idx1 - 1, idx2 - 1, idx3 - 1))
        print('Dump tp {}'.format(wfp))
     
     
    save_prefix = "results/"
    name = face_url.split("/")[-1].split(".")[0] + ".ply"
    print(name)
    face_ply = os.path.join(save_prefix, name)
     
    # 保存 顶点信息 shape 成功
    dump_to_ply(vertices.T, triangles.T, face_ply)   # 切记 tri 是 /Data/uv-data/triangles.txt 中的三角
     
     
    # 保存 带上 color/texture 信息
    def get_colors(image, vertices):
        """
            Args:
                pos: the 3D position map. shape = (256, 256, 3).
            Returns:
                colors: the corresponding colors of vertices. shape = (num of points, 3). n is 45128 here.
        """
        [h, w, _] = image.shape
        vertices[:, 0] = np.minimum(np.maximum(vertices[:, 0], 0), w - 1)  # x
        vertices[:, 1] = np.minimum(np.maximum(vertices[:, 1], 0), h - 1)  # y
        ind = np.round(vertices).astype(np.int32)
        colors = image[ind[:, 1], ind[:, 0], :]  # n x 3
     
        return colors
     
     
    image_face = imread(face_url)  # face_url 是 剪切后为（256， 256， 3）的人脸图像
    [h, w, c] = image_face.shape
    print(h, w, c)
    image_face = image_face / 255.
    colors = get_colors(image_face, vertices)  # 从人脸 和 顶点 中获取 color (43867, 3)
    print(colors.shape)
     
     
    # 写入 .obj文件，具有colors （texture）
    def write_obj_with_colors(obj_name, vertices, triangles, colors):
        ''' Save 3D face model with texture represented by colors.
        Args:
            obj_name: str
            vertices: shape = (nver, 3)
            colors: shape = (nver, 3)
            triangles: shape = (ntri, 3)
        '''
        triangles = triangles.copy()
        triangles += 1  # meshlab start with 1
     
        if obj_name.split('.')[-1] != 'obj':
            obj_name = obj_name + '.obj'
     
        # write obj
        with open(obj_name, 'w') as f:
     
            # write vertices & colors
            for i in range(vertices.shape[0]):
                # s = 'v {} {} {} \n'.format(vertices[0,i], vertices[1,i], vertices[2,i])
                s = 'v {} {} {} {} {} {}\n'.format(vertices[i, 0], vertices[i, 1], vertices[i, 2], colors[i, 0],
                                                   colors[i, 1], colors[i, 2])
                f.write(s)
     
            # write f: ver ind/ uv ind
            [k, ntri] = triangles.shape
            for i in range(triangles.shape[0]):
                # s = 'f {} {} {}\n'.format(triangles[i, 0], triangles[i, 1], triangles[i, 2])
                s = 'f {} {} {}\n'.format(triangles[i, 2], triangles[i, 1], triangles[i, 0])
                f.write(s)
    name = face_url.split("/")[-1].split(".")[0] + ".obj"
    save_vertices = vertices.copy()
    save_vertices[:, 1] = h - 1 - save_vertices[:, 1]  # 这一步 不可缺少； (43867, 3)
    write_obj_with_colors(os.path.join(save_prefix, name), save_vertices, triangles, colors)  # save 3d face(can open with meshlab)

3. obj, ply 文件 用 MeshLab  打开后的效果:

.ply 文件效果：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzIzOTQ0OTE1_size_16_color_FFFFFF_t_70][]

.obj 文件效果：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzIzOTQ0OTE1_size_16_color_FFFFFF_t_70 1][]

从原始UV位置图产生3D人脸，条纹现象不是特别严重， PRNet 预测的效果，条纹效果严重。

代码和文件链接

[https://github.com/DongPoLI/Face-3D-Study/tree/master/generate\_position\_map][https_github.com_DongPoLI_Face-3D-Study_tree_master_generate_position_map]  
————————————————  
版权声明：本文为CSDN博主「挡不住三千问的BlueCat」的原创文章，遵循CC 4.0 BY-SA版权协议，转载请附上原文出处链接及本声明。  
原文链接：https://blog.csdn.net/qq\_23944915/article/details/105046742

[20200323133017663.jpg]: https://img-blog.csdnimg.cn/20200323133017663.jpg
[20200323133042951.jpg]: https://img-blog.csdnimg.cn/20200323133042951.jpg
[20200323133104898.jpg]: /images/20221005/93dffb2d3c7c40d881e4bf40e4f8dbd9.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzIzOTQ0OTE1_size_16_color_FFFFFF_t_70]: /images/20221005/f003c1ba89cc443fa0ffb429ba148867.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzIzOTQ0OTE1_size_16_color_FFFFFF_t_70 1]: /images/20221005/e244e8396346425591f6328b0dfc6bda.png
[https_github.com_DongPoLI_Face-3D-Study_tree_master_generate_position_map]: https://github.com/DongPoLI/Face-3D-Study/tree/master/generate_position_map