nerf三维重建

作者：我家小花儿 | 2024-04-22 05:58:46
踩
nerf三维重建
Nerf（Neural Radiance Fields）是一种用于三维重建和图像合成的机器学习技术。它基于深度学习，使用神经网络来预测场景中每个点的颜色和密度，从而生成高质量的三维重建结果。
Nerf 通过训练神经网络从不同角度的图像中学习场景的表面和光照特征，然后使用学习到的信息来生成新的视角的图像。与传统的三维重建方法不同，Nerf 不需要对场景进行显式的几何建模，也不需要使用多张图像进行立体匹配。相反，它使用单张图像和相机参数来训练神经网络，从而生成高质量的三维重建结果。
Nerf 技术已经被广泛应用于虚拟现实、增强现实和电影等领域，可以生成逼真的三维场景和高质量的图像。同时，它也是当前计算机视觉和深度学习领域的研究热点之一，引起了广泛的关注和研究。
train-nerf.py

import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from tqdm import tqdm
import cv2
import os
import json
import argparse
import imageio

# 主要知识点
# 1. 位置编码，Positional Encoding
#    - 对于输入的x、y、z坐标，因为是连续的无法进行区分，因此采用ff特征，即傅立叶特征进行编码
#    - 编码为cos、sin不同频率的叠加，使得连续值可以具有足够的区分性
# 2. 视图独立性，View Dependent
#    - 输入不仅仅是光线采样点的x、y、z坐标，加上了视图依赖，即x、y、z、theta、pi，5d输入，此时多了射线所在视图
# 3. 分层采样，Hierarchical sampling
#    - 将渲染分为两级，由于第一级别的模型是均匀采样，而实际会有很多无效的采样（即对颜色没有贡献的区域会占比太多），在模型
#       中看来，就是某些点的梯度为0，对模型训练没有贡献
#    - 因此采用两级模型，model、fine，model模型使用均匀采样，推断后得到weights的分布，通过对weights分布进行重采样，使得采样点
#       更加集中在更重要的区域，今儿使得参与训练的点大都是有效的点。所以model作为一级推理，fine则推理重采样后的点
#
# x. 拓展，对于射线的方向和原点的理解，需要具有基本的3d变换知识，建议看GAMES101的前5章补充知识
#    PSNR是峰值信噪比，表示重建的逼真程度
# 这三个环节有了，效果就会非常逼真，但是某些细节上还是存在不足。另外训练时间非常关键

class BlenderProvider:
    def __init__(self, root, transforms_file, half_resolution=True):

        self.meta            = json.load(open(os.path.join(root, transforms_file), "r"))
        self.root            = root
        self.frames          = self.meta["frames"]
        self.images          = []
        self.poses           = []
        self.camera_angle_x  = self.meta["camera_angle_x"]
        
        for frame in self.frames:
            image_file = os.path.join(self.root, frame["file_path"] + ".png")
            image      = imageio.imread(image_file)

            if half_resolution:
                image  = cv2.resize(image, dsize=None, fx=0.5, fy=0.5, interpolation=cv2.INTER_AREA)

            self.images.append(image)
            self.poses.append(frame["transform_matrix"])

        self.poses  = np.stack(self.poses)
        self.images = (np.stack(self.images) / 255.0).astype(np.float32)
        self.width  = self.images.shape[2]
        self.height = self.images.shape[1]
        self.focal  = 0.5 * self.width / np.tan(0.5 * self.camera_angle_x)

        alpha       = self.images[..., [3]]
        rgb         = self.images[..., :3]
        self.images = rgb * alpha + (1 - alpha)


class NeRFDataset:
    def __init__(self, provider, batch_size=1024, device="cpu"):

        self.images        = provider.images
        self.poses         = provider.poses
        self.focal         = provider.focal
        self.width         = provider.width
        self.height        = provider.height
        self.batch_size    = batch_size
        self.num_image     = len(self.images)
        self.precrop_iters = 500
        self.precrop_frac  = 0.5
        self.niter         = 0
        self.device        = device
        self.initialize()


    def initialize(self):

        warange = torch.arange(self.width,  dtype=torch.float32, device=self.device)
        harange = torch.arange(self.height, dtype=torch.float32, device=self.device)
        y, x = torch.meshgrid(harange, warange)

        self.transformed_x = (x - self.width * 0.5) / self.focal
        self.transformed_y = (y - self.height * 0.5) / self.focal

        # pre center crop
        self.precrop_index = torch.arange(self.width * self.height).view(self.height, self.width)

        dH = int(self.height // 2 * self.precrop_frac)
        dW = int(self.width  // 2 * self.precrop_frac)
        self.precrop_index = self.precrop_index[
            self.height // 2 - dH:self.height // 2 + dH, 
            self.width  // 2 - dW:self.width  // 2 + dW
        ].reshape(-1)

        poses = torch.FloatTensor(self.poses, device=self.device)
        all_ray_dirs, all_ray_origins = [], []

        for i in range(len(self.images)):
            ray_dirs, ray_origins = self.make_rays(self.transformed_x, self.transformed_y, poses[i])
            all_ray_dirs.append(ray_dirs)
            all_ray_origins.append(ray_origins)

        self.all_ray_dirs    = torch.stack(all_ray_dirs, dim=0)
        self.all_ray_origins = torch.stack(all_ray_origins, dim=0)
        self.images          = torch.FloatTensor(self.images, device=self.device).view(self.num_image, -1, 3)
        

    def __getitem__(self, index):
        self.niter += 1

        ray_dirs   = self.all_ray_dirs[index]
        ray_oris   = self.all_ray_origins[index]
        img_pixels = self.images[index]
        if self.niter < self.precrop_iters:
            ray_dirs   = ray_dirs[self.precrop_index]
            ray_oris   = ray_oris[self.precrop_index]
            img_pixels = img_pixels[self.precrop_index]

        nrays          = self.batch_size
        select_inds    = np.random.choice(ray_dirs.shape[0], size=[nrays], replace=False)
        ray_dirs       = ray_dirs[select_inds]
        ray_oris       = ray_oris[select_inds]
        img_pixels     = img_pixels[select_inds]

        # dirs是指：direction
        # ori是指： origin
        return ray_dirs, ray_oris, img_pixels


    def __len__(self):
        return self.num_image


    def make_rays(self, x, y, pose):

        # 100, 100, 3
        # 坐标系在-y，-z方向上
        directions    = torch.stack([x, -y, -torch.ones_like(x)], dim=-1)
        camera_matrix = pose[:3, :3]
        
        # 10000 x 3
        ray_dirs = directions.reshape(-1, 3) @ camera_matrix.T
        ray_origin = pose[:3, 3].view(1, 3).repeat(len(ray_dirs), 1)
        return ray_dirs, ray_origin


    def get_test_item(self, index=0):

        ray_dirs   = self.all_ray_dirs[index]
        ray_oris   = self.all_ray_origins[index]
        img_pixels = self.images[index]

        for i in range(0, len(ray_dirs), self.batch_size):
            yield ray_dirs[i:i+self.batch_size], ray_oris[i:i+self.batch_size], img_pixels[i:i+self.batch_size]


    def get_rotate_360_rays(self):
        def trans_t(t):
            return np.array([
                [1,0,0,0],
                [0,1,0,0],
                [0,0,1,t],
                [0,0,0,1],
            ], dtype=np.float32)

        def rot_phi(phi):
            return np.array([
                [1,0,0,0],
                [0,np.cos(phi),-np.sin(phi),0],
                [0,np.sin(phi), np.cos(phi),0],
                [0,0,0,1],
            ], dtype=np.float32)

        def rot_theta(th) : 
            return np.array([
                [np.cos(th),0,-np.sin(th),0],
                [0,1,0,0],
                [np.sin(th),0, np.cos(th),0],
                [0,0,0,1],
            ], dtype=np.float32)

        def pose_spherical(theta, phi, radius):
            c2w = trans_t(radius)
            c2w = rot_phi(phi/180.*np.pi) @ c2w
            c2w = rot_theta(theta/180.*np.pi) @ c2w
            c2w = np.array([[-1,0,0,0],[0,0,1,0],[0,1,0,0],[0,0,0,1]]) @ c2w
            return c2w

        for th in np.linspace(-180., 180., 41, endpoint=False):
            pose = torch.FloatTensor(pose_spherical(th, -30., 4.), device=self.device)

            def genfunc():
                ray_dirs, ray_origins = self.make_rays(self.transformed_x, self.transformed_y, pose)
                for i in range(0, len(ray_dirs), 1024):
                    yield ray_dirs[i:i+1024], ray_origins[i:i+1024]

            yield genfunc


# Hierarchical sampling (section 5.2)
def sample_pdf(bins, weights, N_samples, det=False):
    # Get pdf
    device = weights.device
    weights = weights + 1e-5 # prevent nans
    pdf = weights / torch.sum(weights, -1, keepdim=True)
    cdf = torch.cumsum(pdf, -1)
    cdf = torch.cat([torch.zeros_like(cdf[...,:1]), cdf], -1)  # (batch, len(bins))
    
    # Take uniform samples
    if det:
        u = torch.linspace(0., 1., steps=N_samples, device=device)
        u = u.expand(list(cdf.shape[:-1]) + [N_samples])
    else:
        u = torch.rand(list(cdf.shape[:-1]) + [N_samples])

    # Invert CDF
    u = u.contiguous()
    inds = torch.searchsorted(cdf, u, right=True)
    below = torch.max(torch.zeros_like(inds-1), inds-1)
    above = torch.min((cdf.shape[-1]-1) * torch.ones_like(inds), inds)
    inds_g = torch.stack([below, above], -1)  # (batch, N_samples, 2)

    # cdf_g = tf.gather(cdf, inds_g, axis=-1, batch_dims=len(inds_g.shape)-2)
    # bins_g = tf.gather(bins, inds_g, axis=-1, batch_dims=len(inds_g.shape)-2)
    matched_shape = [inds_g.shape[0], inds_g.shape[1], cdf.shape[-1]]
    cdf_g = torch.gather(cdf.unsqueeze(1).expand(matched_shape), 2, inds_g)
    bins_g = torch.gather(bins.unsqueeze(1).expand(matched_shape), 2, inds_g)

    denom = (cdf_g[...,1]-cdf_g[...,0])
    denom = torch.where(denom<1e-5, torch.ones_like(denom), denom)
    t = (u-cdf_g[...,0])/denom
    samples = bins_g[...,0] + t * (bins_g[...,1]-bins_g[...,0])
    return samples


def sample_rays(ray_directions, ray_origins, sample_z_vals):

    nrays = len(ray_origins)
    sample_z_vals = sample_z_vals.repeat(nrays, 1)
    # a.shape = (3, 3, 4, 5)  a[..., 4]
    rays = ray_origins[:, None, :] + ray_directions[:, None, :] * sample_z_vals[..., None]
    return rays, sample_z_vals


def sample_viewdirs(ray_directions):
    return ray_directions / torch.norm(ray_directions, dim=-1, keepdim=True)
    

def predict_to_rgb(sigma, rgb, z_vals, raydirs, white_background=False):

    device         = sigma.device
    delta_prefix   = z_vals[..., 1:] - z_vals[..., :-1]
    delta_addition = torch.full((z_vals.size(0), 1), 1e10, device=device)
    delta = torch.cat([delta_prefix, delta_addition], dim=-1)
    delta = delta * torch.norm(raydirs[..., None, :], dim=-1)

    alpha    = 1.0 - torch.exp(-sigma * delta)
    exp_term = 1.0 - alpha
    epsilon  = 1e-10
    exp_addition = torch.ones(exp_term.size(0), 1, device=device)
    exp_term = torch.cat([exp_addition, exp_term + epsilon], dim=-1)
    transmittance = torch.cumprod(exp_term, axis=-1)[..., :-1]

    weights       = alpha * transmittance
    rgb           = torch.sum(weights[..., None] * rgb, dim=-2)
    depth         = torch.sum(weights * z_vals, dim=-1)
    acc_map       = torch.sum(weights, -1)

    if white_background:
        rgb       = rgb + (1.0 - acc_map[..., None])
    return rgb, depth, acc_map, weights


def render_rays(model, fine, raydirs, rayoris, sample_z_vals, importance=0, white_background=False):

    rays, z_vals = sample_rays(raydirs, rayoris, sample_z_vals)
    view_dirs    = sample_viewdirs(raydirs)

    sigma, rgb = model(rays, view_dirs)
    sigma      = sigma.squeeze(dim=-1)
    rgb1, depth1, acc_map1, weights1 = predict_to_rgb(sigma, rgb, z_vals, raydirs, white_background)

    # 使用weights1进行重采样
    z_vals_mid = .5 * (z_vals[...,1:] + z_vals[...,:-1])
    z_samples  = sample_pdf(z_vals_mid, weights1[...,1:-1], importance, det=True)
    z_samples  = z_samples.detach()

    z_vals, _  = torch.sort(torch.cat([z_vals, z_samples], -1), -1)
    rays       = rayoris[...,None,:] + raydirs[...,None,:] * z_vals[...,:,None] # [N_rays, N_samples + N_importance, 3]
    sigma, rgb = fine(rays, view_dirs)
    sigma      = sigma.squeeze(dim=-1)

    # 第二次重采样的预测才是最终结果，这是论文中，分层采样环节（Hierarchical sampling）
    rgb2, depth2, acc_map2, weights2 = predict_to_rgb(sigma, rgb, z_vals, raydirs, white_background)
    return rgb1, rgb2

# 无视图独立性的head
class NoViewDirHead(nn.Module):
    def __init__(self, ninput, noutput):
        super().__init__()

        self.head = nn.Linear(ninput, noutput)
    
    def forward(self, x, view_dirs):
        
        x = self.head(x)
        rgb   = x[..., :3].sigmoid()
        sigma = x[..., 3].relu()
        return sigma, rgb

# 视图独立性的head
class ViewDenepdentHead(nn.Module):
    def __init__(self, ninput, nview):
        super().__init__()

        self.feature = nn.Linear(ninput, ninput)
        self.view_fc = nn.Linear(ninput + nview, ninput // 2)
        self.alpha = nn.Linear(ninput, 1)
        self.rgb = nn.Linear(ninput // 2, 3)
    
    def forward(self, x, view_dirs):
        
        feature = self.feature(x)
        sigma   = self.alpha(x).relu()
        feature = torch.cat([feature, view_dirs], dim=-1)
        feature = self.view_fc(feature).relu()
        rgb     = self.rgb(feature).sigmoid()
        return sigma, rgb

# 位置编码实现
class Embedder(nn.Module):
    def __init__(self, positional_encoding_dim):
        super().__init__()
        self.positional_encoding_dim = positional_encoding_dim

    def forward(self, x):
        
        positions = [x]
        for i in range(self.positional_encoding_dim):
            for fn in [torch.sin, torch.cos]:
                positions.append(fn((2.0 ** i) * x))

        return torch.cat(positions, dim=-1)

# 基本模型结构
class NeRF(nn.Module):
    def __init__(self, x_pedim=10, nwidth=256, ndepth=8, view_pedim=4):
        super().__init__()
        
        xdim         = (x_pedim * 2 + 1) * 3

        layers       =  []
        layers_in    = [nwidth] * ndepth
        layers_in[0] = xdim
        layers_in[5] = nwidth + xdim

        # 模型中特定层[5]会存在concat
        for i in range(ndepth):
            layers.append(nn.Linear(layers_in[i], nwidth))
        
        if view_pedim > 0:
            view_dim = (view_pedim * 2 + 1) * 3
            self.view_embed = Embedder(view_pedim)
            self.head = ViewDenepdentHead(nwidth, view_dim)
        else:
            self.head = NoViewDirHead(nwidth, 4)
        
        self.xembed = Embedder(x_pedim)
        self.layers = nn.Sequential(*layers)

    
    def forward(self, x, view_dirs):
        
        xshape = x.shape
        x = self.xembed(x)
        if self.view_embed is not None:
            view_dirs = view_dirs[:, None].expand(xshape)
            view_dirs = self.view_embed(view_dirs)

        raw_x = x
        for i, layer in enumerate(self.layers):
            x = torch.relu(layer(x))
            
            if i == 4:
                x = torch.cat([x, raw_x], axis=-1)

        return self.head(x, view_dirs)


def train():

    pbar     = tqdm(range(1, maxiters))
    for global_step in pbar:

        idx   = np.random.randint(0, len(trainset))
        raydirs, rayoris, imagepixels = trainset[idx]

        rgb1, rgb2 = render_rays(model, fine, raydirs, rayoris, sample_z_vals, importance, white_background)
        loss1 = ((rgb1 - imagepixels)**2).mean()
        loss2 = ((rgb2 - imagepixels)**2).mean()
        psnr  = -10. * torch.log(loss2.detach()) / np.log(10.)
        loss  = loss1 + loss2
        
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        pbar.set_description(f"{global_step} / {maxiters}, Loss: {loss.item():.6f}, PSNR: {psnr.item():.6f}")

        decay_rate = 0.1
        new_lrate  = lrate * (decay_rate ** (global_step / lrate_decay))
        for param_group in optimizer.param_groups:
            param_group['lr'] = new_lrate

        if global_step % 5000 == 0 or global_step == 500:

            imgpath = f"imgs/{global_step:02d}.png"
            pthpath = f"ckpt/{global_step:02d}.pth"
            model.eval()
            with torch.no_grad():
                rgbs, imgpixels = [], []
                for raydirs, rayoris, imagepixels in trainset.get_test_item():

                    rgb1, rgb2  = render_rays(model, fine, raydirs, rayoris, sample_z_vals, importance, white_background)
                    rgbs.append(rgb2)
                    imgpixels.append(imagepixels)

                rgb       = torch.cat(rgbs, dim=0)
                imgpixels = torch.cat(imgpixels, dim=0)
                loss      = ((rgb - imgpixels)**2).mean()
                psnr      = -10. * torch.log(loss) / np.log(10.)

                print(f"Save image {imgpath}, Loss: {loss.item():.6f}, PSNR: {psnr.item():.6f}")
            model.train()
            
            temp_image = (rgb.view(height, width, 3).cpu().numpy() * 255).astype(np.uint8)
            cv2.imwrite(imgpath, temp_image[..., ::-1])
            torch.save([model.state_dict(), fine.state_dict()], pthpath)


def make_video360():

    mstate, fstate = torch.load(args.ckpt, map_location="cpu")
    model.load_state_dict(mstate)
    fine.load_state_dict(fstate)
    model.eval()
    fine.eval()
    imagelist = []

    for i, gfn in tqdm(enumerate(trainset.get_rotate_360_rays()), desc="Rendering"):

        with torch.no_grad():
            rgbs = []
            for raydirs, rayoris in gfn():
                rgb1, rgb2 = render_rays(model, fine, raydirs, rayoris, sample_z_vals, importance, white_background)
                rgbs.append(rgb2)

            rgb = torch.cat(rgbs, dim=0)
        
        rgb  = (rgb.view(height, width, 3).cpu().numpy() * 255).astype(np.uint8)
        file = f"rotate360/{i:03d}.png"

        print(f"Rendering to {file}")
        cv2.imwrite(file, rgb[..., ::-1])
        imagelist.append(rgb)

    video_file = f"videos/rotate360.mp4"
    print(f"Write imagelist to video file {video_file}")
    imageio.mimwrite(video_file, imagelist, fps=30, quality=10)


if __name__ == "__main__":

    parser = argparse.ArgumentParser()
    parser.add_argument("--datadir", type=str, default='data/nerf_synthetic/lego', help='input data directory')
    parser.add_argument("--make-video360", action="store_true", help="make 360 rotation video")
    parser.add_argument("--half-resolution", action="store_true", help="use half resolution")
    parser.add_argument("--ckpt", default="300000.pth", type=str, help="model file used to make 360 rotation video")
    args = parser.parse_args()

    device      = "cpu"
    maxiters    = 100000 + 1
    batch_size  = 1024
    lrate_decay = 500 * 1000
    lrate       = 5e-4
    importance  = 128
    num_samples = 64                    # 每个光线的采样数量
    positional_encoding_dim = 10        # 位置编码维度
    view_encoding_dim       = 4         # View Dependent对应的位置编码维度
    white_background        = True      # 图片背景是白色的
    half_resolution         = args.half_resolution    # 只进行一半分辨率的重建(400x400)，False表示(800x800)分辨率
    sample_z_vals           = torch.linspace(2.0, 6.0, num_samples, device=device).view(1, num_samples)

    model = NeRF(
        x_pedim    = positional_encoding_dim,
        view_pedim = view_encoding_dim
    ).to(device)
    params = list(model.parameters())
    
    # 使用model产生的权重进行重采样，然后再推理，所以这个才是效果更好的模型
    fine = NeRF(
        x_pedim    = positional_encoding_dim,
        view_pedim = view_encoding_dim
    ).to(device)
    params.extend(list(fine.parameters()))

    optimizer = optim.Adam(params, lrate)
    os.makedirs("imgs",      exist_ok=True)
    os.makedirs("rotate360", exist_ok=True)
    os.makedirs("videos",    exist_ok=True)
    os.makedirs("ckpt",      exist_ok=True)

    print(model)

    provider = BlenderProvider("data/nerf_synthetic/lego", "transforms_train.json", half_resolution)
    trainset = NeRFDataset(provider, batch_size, device)
    width    = trainset.width
    height   = trainset.height

    if args.make_video360:
        make_video360()
    else:
        train()

    print("Program done.")
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
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524