train.py:
def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoint_iterations, checkpoint, debug_from):
first_iter = 0
tb_writer = prepare_output_and_logger(dataset)
gaussians = GaussianModel(dataset.sh_degree) # 创建GaussianModel模型,给点云中的每个点去创建一个3D gaussian
scene = Scene(dataset, gaussians) # 加载数据集和每张图片对应的camera的参数
gaussians.training_setup(opt)
if checkpoint: #如果checkpoint非空,则加载checkpoint,恢复模型参数和优化参数
(model_params, first_iter) = torch.load(checkpoint)
gaussians.restore(model_params, opt)
#设置背景值,白色[1, 1, 1],否则[0, 0, 0]
bg_color = [1, 1, 1] if dataset.white_background else [0, 0, 0]
background = torch.tensor(bg_color, dtype=torch.float32, device="cuda")
#记录gpu运行时间
iter_start = torch.cuda.Event(enable_timing = True)
iter_end = torch.cuda.Event(enable_timing = True)
####################迭代训练开始###########################
viewpoint_stack = None
ema_loss_for_log = 0.0
progress_bar = tqdm(range(first_iter, opt.iterations), desc="Training progress")
first_iter += 1
for iteration in range(first_iter, opt.iterations + 1):
if network_gui.conn == None:
network_gui.try_connect()
while network_gui.conn != None:
try:
net_image_bytes = None
custom_cam, do_training, pipe.convert_SHs_python, pipe.compute_cov3D_python, keep_alive, scaling_modifer = network_gui.receive()
if custom_cam != None:
net_image = render(custom_cam, gaussians, pipe, background, scaling_modifer)["render"]
net_image_bytes = memoryview((torch.clamp(net_image, min=0, max=1.0) * 255).byte().permute(1, 2, 0).contiguous().cpu().numpy())
network_gui.send(net_image_bytes, dataset.source_path)
if do_training and ((iteration < int(opt.iterations)) or not keep_alive):
break
except Exception as e:
network_gui.conn = None
#记录开始时间
iter_start.record()
# 对xyz的学习率进行调整
gaussians.update_learning_rate(iteration)
# Every 1000 its we increase the levels of SH up to a maximum degree
if iteration % 1000 == 0:
gaussians.oneupSHdegree()
# Pick a random Camera #随机选择一个图片及其相应的相机视角(内外参)
if not viewpoint_stack:
viewpoint_stack = scene.getTrainCameras().copy()
viewpoint_cam = viewpoint_stack.pop(randint(0, len(viewpoint_stack)-1))
# Render,在debug_from步时,启动调试
if (iteration - 1) == debug_from:
pipe.debug = True
#随机背景与否
bg = torch.rand((3), device="cuda") if opt.random_background else background
#render 函数参数一:上步随机选择的相机,参数二:高斯模型,参数三:管道参数,参数四:背景
render_pkg = render(viewpoint_cam, gaussians, pipe, bg)
image, viewspace_point_tensor, visibility_filter, radii = render_pkg["render"], render_pkg["viewspace_points"], render_pkg["visibility_filter"], render_pkg["radii"]
# Loss
gt_image = viewpoint_cam.original_image.cuda()
Ll1 = l1_loss(image, gt_image)
loss = (1.0 - opt.lambda_dssim) * Ll1 + opt.lambda_dssim * (1.0 - ssim(image, gt_image))
loss.backward()
#记录结束时间
iter_end.record()
with torch.no_grad():
# Progress bar更新进度条,每10步更新一下后缀内容及进度
ema_loss_for_log = 0.4 * loss.item() + 0.6 * ema_loss_for_log
if iteration % 10 == 0:
progress_bar.set_postfix({"Loss": f"{ema_loss_for_log:.{7}f}"})
progress_bar.update(10)
if iteration == opt.iterations:
progress_bar.close()
# Log and save
training_report(tb_writer, iteration, Ll1, loss, l1_loss, iter_start.elapsed_time(iter_end), testing_iterations, scene, render, (pipe, background))
if (iteration in saving_iterations):
print("\n[ITER {}] Saving Gaussians".format(iteration))
scene.save(iteration)
# Densification
if iteration < opt.densify_until_iter:#15000步以前
# Keep track of max radii in image-space for pruning
gaussians.max_radii2D[visibility_filter] = torch.max(gaussians.max_radii2D[visibility_filter], radii[visibility_filter])
# 统计3D gaussian均值(xyz)的梯度, 用于对3D gaussians的克隆或者切分
gaussians.add_densification_stats(viewspace_point_tensor, visibility_filter)
#每100步进行下面的操作,对3D gaussians进行克隆或者切分, 并将opacity小于一定阈值的3D gaussians进行删除
if iteration > opt.densify_from_iter and iteration % opt.densification_interval == 0:
size_threshold = 20 if iteration > opt.opacity_reset_interval else None
gaussians.densify_and_prune(opt.densify_grad_threshold, 0.005, scene.cameras_extent, size_threshold)
#每3000步重设不透明度
if iteration % opt.opacity_reset_interval == 0 or (dataset.white_background and iteration == opt.densify_from_iter):
gaussians.reset_opacity()
# Optimizer step
if iteration < opt.iterations:#30000步以前
gaussians.optimizer.step()
gaussians.optimizer.zero_grad(set_to_none = True)
if (iteration in checkpoint_iterations):
print("\n[ITER {}] Saving Checkpoint".format(iteration))
torch.save((gaussians.capture(), iteration), scene.model_path + "/chkpnt" + str(iteration) + ".pth")
####################迭代训练结束###########################
def prepare_output_and_logger(args):
if not args.model_path:
if os.getenv('OAR_JOB_ID'):
unique_str=os.getenv('OAR_JOB_ID')
else:
unique_str = str(uuid.uuid4())
#生成一个随机的UUID,并取前十位作为输出目录名称
args.model_path = os.path.join("./output/", unique_str[0:10])
# Set up output folder
print("Output folder: {}".format(args.model_path))
os.makedirs(args.model_path, exist_ok = True)
#配置参数写入到cfg_args文件中
with open(os.path.join(args.model_path, "cfg_args"), 'w') as cfg_log_f:
cfg_log_f.write(str(Namespace(**vars(args))))
# Create Tensorboard writer
tb_writer = None
if TENSORBOARD_FOUND:
tb_writer = SummaryWriter(args.model_path)
else:
print("Tensorboard not available: not logging progress")
return tb_writer
#写入日志
def training_report(tb_writer, iteration, Ll1, loss, l1_loss, elapsed, testing_iterations, scene : Scene, renderFunc, renderArgs):
if tb_writer:
tb_writer.add_scalar('train_loss_patches/l1_loss', Ll1.item(), iteration)
tb_writer.add_scalar('train_loss_patches/total_loss', loss.item(), iteration)
tb_writer.add_scalar('iter_time', elapsed, iteration)
# Report test and samples of training set
if iteration in testing_iterations:
torch.cuda.empty_cache()
validation_configs = ({'name': 'test', 'cameras' : scene.getTestCameras()},
{'name': 'train', 'cameras' : [scene.getTrainCameras()[idx % len(scene.getTrainCameras())] for idx in range(5, 30, 5)]})
for config in validation_configs:
if config['cameras'] and len(config['cameras']) > 0:
l1_test = 0.0
psnr_test = 0.0
for idx, viewpoint in enumerate(config['cameras']):
image = torch.clamp(renderFunc(viewpoint, scene.gaussians, *renderArgs)["render"], 0.0, 1.0)
gt_image = torch.clamp(viewpoint.original_image.to("cuda"), 0.0, 1.0)
if tb_writer and (idx < 5):
tb_writer.add_images(config['name'] + "_view_{}/render".format(viewpoint.image_name), image[None], global_step=iteration)
if iteration == testing_iterations[0]:
tb_writer.add_images(config['name'] + "_view_{}/ground_truth".format(viewpoint.image_name), gt_image[None], global_step=iteration)
l1_test += l1_loss(image, gt_image).mean().double()
psnr_test += psnr(image, gt_image).mean().double()
psnr_test /= len(config['cameras'])
l1_test /= len(config['cameras'])
print("\n[ITER {}] Evaluating {}: L1 {} PSNR {}".format(iteration, config['name'], l1_test, psnr_test))
if tb_writer:
tb_writer.add_scalar(config['name'] + '/loss_viewpoint - l1_loss', l1_test, iteration)
tb_writer.add_scalar(config['name'] + '/loss_viewpoint - psnr', psnr_test, iteration)
if tb_writer:
tb_writer.add_histogram("scene/opacity_histogram", scene.gaussians.get_opacity, iteration)
tb_writer.add_scalar('total_points', scene.gaussians.get_xyz.shape[0], iteration)
torch.cuda.empty_cache()
if __name__ == "__main__":
# Set up command line argument parser
parser = ArgumentParser(description="Training script parameters")
lp = ModelParams(parser)
op = OptimizationParams(parser)
pp = PipelineParams(parser)
parser.add_argument('--ip', type=str, default="127.0.0.1")
parser.add_argument('--port', type=int, default=6009)
parser.add_argument('--debug_from', type=int, default=-1)
parser.add_argument('--detect_anomaly', action='store_true', default=False)
parser.add_argument("--test_iterations", nargs="+", type=int, default=[7_000, 30_000])
parser.add_argument("--save_iterations", nargs="+", type=int, default=[7_000, 30_000])
parser.add_argument("--quiet", action="store_true")
parser.add_argument("--checkpoint_iterations", nargs="+", type=int, default=[])
parser.add_argument("--start_checkpoint", type=str, default = None)
args = parser.parse_args(sys.argv[1:])
args.save_iterations.append(args.iterations)
print("Optimizing " + args.model_path)
# Initialize system state (RNG 随机数生成器)
safe_state(args.quiet)
# Start GUI server, configure and run training
# 在ip的port端口监听
network_gui.init(args.ip, args.port)
# 正向传播时:关闭自动求导的异常侦测
torch.autograd.set_detect_anomaly(args.detect_anomaly)
#开始训练,函数参数:模型参数lp,优化参数,管道参数,测试迭代值,保存迭代值,检查点迭代值,开始检查点,调试来源。
training(lp.extract(args), op.extract(args), pp.extract(args), args.test_iterations, args.save_iterations, args.checkpoint_iterations, args.start_checkpoint, args.debug_from)
#结束
# All done
print("\nTraining complete.")
scene/__init__.py
class Scene:
gaussians : GaussianModel
#场景初始化,参数:模型参数和高斯模型
def __init__(self, args : ModelParams, gaussians : GaussianModel, load_iteration=None, shuffle=True, resolution_scales=[1.0]):
"""b
:param path: Path to colmap scene main folder.
"""
self.model_path = args.model_path
self.loaded_iter = None
self.gaussians = gaussians
# 是否加载迭代,load_iteration是迭代值
# 寻找是否有训练过的记录, 如果没有则为初次训练, 需要从COLMAP创建的点云中初始化每个点对应的3D gaussian
# 以及将每张图片对应的相机参数dump到`cameras.json`文件中
if load_iteration:
if load_iteration == -1:
self.loaded_iter = searchForMaxIteration(os.path.join(self.model_path, "point_cloud"))
else:
self.loaded_iter = load_iteration
print("Loading trained model at iteration {}".format(self.loaded_iter))
self.train_cameras = {}
self.test_cameras = {}
# 从COLMAP或Blender中读取每张图片, 以及每张图片对应的相机内外参
# 如果源目录中有sparse目录,则读取数据得到场景信息scene_info
if os.path.exists(os.path.join(args.source_path, "sparse")):
scene_info = sceneLoadTypeCallbacks["Colmap"](args.source_path, args.images, args.eval)
#如果有transforms_train.json文件,则调用readNerfSyntheticInfo
elif os.path.exists(os.path.join(args.source_path, "transforms_train.json")):
print("Found transforms_train.json file, assuming Blender data set!")
scene_info = sceneLoadTypeCallbacks["Blender"](args.source_path, args.white_background, args.eval)
else:
assert False, "Could not recognize scene type!"
# 如果loaded_iter为空(没有设置从已经训练过的位置开始训练)
# 将每张图片对应的相机参数dump到`cameras.json`文件中
if not self.loaded_iter:
#拷贝ply文件到输出目录,命名为input.ply
with open(scene_info.ply_path, 'rb') as src_file, open(os.path.join(self.model_path, "input.ply") , 'wb') as dest_file:
dest_file.write(src_file.read())
json_cams = []
#camlist追加scene_info的测试集和训练集
camlist = []
if scene_info.test_cameras:
camlist.extend(scene_info.test_cameras)
if scene_info.train_cameras:
camlist.extend(scene_info.train_cameras)
#把camlist的数据转成json存到cameras.json文件中
for id, cam in enumerate(camlist):
json_cams.append(camera_to_JSON(id, cam))
with open(os.path.join(self.model_path, "cameras.json"), 'w') as file:
json.dump(json_cams, file)
# 随机打乱所有图片和对应相机的顺序,打乱train_cameras
if shuffle:
random.shuffle(scene_info.train_cameras) # Multi-res consistent random shuffling
random.shuffle(scene_info.test_cameras) # Multi-res consistent random shuffling
# 把getNerfppNorm返回的结果的半径赋给cameras_extent,所有相机的中心点位置到最远camera的距离
self.cameras_extent = scene_info.nerf_normalization["radius"]
#对每一个分辨率缩放,计算cameraList
for resolution_scale in resolution_scales:
print("Loading Training Cameras")
self.train_cameras[resolution_scale] = cameraList_from_camInfos(scene_info.train_cameras, resolution_scale, args)
print("Loading Test Cameras")
self.test_cameras[resolution_scale] = cameraList_from_camInfos(scene_info.test_cameras, resolution_scale, args)
# 如果是初次训练, 则从COLMAP创建的点云中初始化每个点对应的3D gaussian, 否则直接从之前保存的模型文件中读取3D gaussian
if self.loaded_iter:
self.gaussians.load_ply(os.path.join(self.model_path,
"point_cloud",
"iteration_" + str(self.loaded_iter),
"point_cloud.ply"))
else:
self.gaussians.create_from_pcd(scene_info.point_cloud, self.cameras_extent)
scene/gaussian_model.py
class GaussianModel:
def setup_functions(self):
# 从尺度和旋转参数中去构建3Dgaussian的协方差矩阵,L=RS,L@L的逆等于RSStRt,再去除对称性得到6个值表示对称的协方差矩阵
def build_covariance_from_scaling_rotation(scaling, scaling_modifier, rotation):
L = build_scaling_rotation(scaling_modifier * scaling, rotation)
actual_covariance = L @ L.transpose(1, 2)
symm = strip_symmetric(actual_covariance)
return symm
self.scaling_activation = torch.exp # 将尺度限制为非负数
self.scaling_inverse_activation = torch.log
self.covariance_activation = build_covariance_from_scaling_rotation
self.opacity_activation = torch.sigmoid # 将不透明度限制在0-1的范围内
self.inverse_opacity_activation = inverse_sigmoid
self.rotation_activation = torch.nn.functional.normalize
def __init__(self, sh_degree : int):
self.active_sh_degree = 0
self.max_sh_degree = sh_degree
self._xyz = torch.empty(0) # 中心点位置, 也即3Dgaussian的均值
self._features_dc = torch.empty(0) # 第一个球谐系数, 球谐系数用来表示RGB颜色
self._features_rest = torch.empty(0) # 其余球谐系数
self._scaling = torch.empty(0) # 尺度
self._rotation = torch.empty(0) # 旋转参数, 四元组
self._opacity = torch.empty(0) # 不透明度
self.max_radii2D = torch.empty(0) # 投影到2D时, 每个2D gaussian最大的半径
self.xyz_gradient_accum = torch.empty(0) # 3Dgaussian的均值的累积梯度
self.denom = torch.empty(0)
self.optimizer = None
self.percent_dense = 0
self.spatial_lr_scale = 0
self.setup_functions()
def capture(self):
return (
self.active_sh_degree,
self._xyz,
self._features_dc,
self._features_rest,
self._scaling,
self._rotation,
self._opacity,
self.max_radii2D,
self.xyz_gradient_accum,
self.denom,
self.optimizer.state_dict(),
self.spatial_lr_scale,
)
def restore(self, model_args, training_args):
(self.active_sh_degree,
self._xyz,
self._features_dc,
self._features_rest,
self._scaling,
self._rotation,
self._opacity,
self.max_radii2D,
xyz_gradient_accum,
denom,
opt_dict,
self.spatial_lr_scale) = model_args
self.training_setup(training_args)
self.xyz_gradient_accum = xyz_gradient_accum
self.denom = denom
self.optimizer.load_state_dict(opt_dict)
@property
def get_scaling(self):
return self.scaling_activation(self._scaling)
@property
def get_rotation(self):
return self.rotation_activation(self._rotation)
@property
def get_xyz(self):
return self._xyz
@property
def get_features(self):
features_dc = self._features_dc
features_rest = self._features_rest
return torch.cat((features_dc, features_rest), dim=1)
@property
def get_opacity(self):
return self.opacity_activation(self._opacity)
def get_covariance(self, scaling_modifier = 1):
return self.covariance_activation(self.get_scaling, scaling_modifier, self._rotation)
def oneupSHdegree(self):
if self.active_sh_degree < self.max_sh_degree:
self.active_sh_degree += 1
#参数一:稀疏点云,参数二:球半径
def create_from_pcd(self, pcd : BasicPointCloud, spatial_lr_scale : float):
self.spatial_lr_scale = spatial_lr_scale
#点云转tensor类型,送到gpu
fused_point_cloud = torch.tensor(np.asarray(pcd.points)).float().cuda()
#把点云的颜色转成sh系数,送到gpu
fused_color = RGB2SH(torch.tensor(np.asarray(pcd.colors)).float().cuda())
#特征长度为16,第一个维度为sh系数,其他赋0
features = torch.zeros((fused_color.shape[0], 3, (self.max_sh_degree + 1) ** 2)).float().cuda()
features[:, :3, 0 ] = fused_color
features[:, 3:, 1:] = 0.0
print("Number of points at initialisation : ", fused_point_cloud.shape[0])
#调用simple_knn的distCUDA2函数,计算点云中的每个点到与其最近的K个点的平均距离的平方
dist2 = torch.clamp_min(distCUDA2(torch.from_numpy(np.asarray(pcd.points)).float().cuda()), 0.0000001)
#距离值取对数,得到缩放值,repeat变成3个
scales = torch.log(torch.sqrt(dist2))[...,None].repeat(1, 3)
rots = torch.zeros((fused_point_cloud.shape[0], 4), device="cuda")
rots[:, 0] = 1
#0.1的逆sigmoid值-2.1976,也即是-2.1976的sigmoid值0.1,初始化设置为不透明度
opacities = inverse_sigmoid(0.1 * torch.ones((fused_point_cloud.shape[0], 1), dtype=torch.float, device="cuda"))
#初始化位置,sh系数(直接+剩余),缩放(3个轴),旋转(四元数),不透明度(逆sigmoid的值),
self._xyz = nn.Parameter(fused_point_cloud.requires_grad_(True))
self._features_dc = nn.Parameter(features[:,:,0:1].transpose(1, 2).contiguous().requires_grad_(True))
self._features_rest = nn.Parameter(features[:,:,1:].transpose(1, 2).contiguous().requires_grad_(True))
self._scaling = nn.Parameter(scales.requires_grad_(True))
self._rotation = nn.Parameter(rots.requires_grad_(True))
self._opacity = nn.Parameter(opacities.requires_grad_(True))
self.max_radii2D = torch.zeros((self.get_xyz.shape[0]), device="cuda")
#
def training_setup(self, training_args):
self.percent_dense = training_args.percent_dense
self.xyz_gradient_accum = torch.zeros((self.get_xyz.shape[0], 1), device="cuda")
self.denom = torch.zeros((self.get_xyz.shape[0], 1), device="cuda")
#参数列表,每一项是一个字典,包含位置、f_dc、f_rest、不透明度、缩放、旋转
l = [
{'params': [self._xyz], 'lr': training_args.position_lr_init * self.spatial_lr_scale, "name": "xyz"},
{'params': [self._features_dc], 'lr': training_args.feature_lr, "name": "f_dc"},
{'params': [self._features_rest], 'lr': training_args.feature_lr / 20.0, "name": "f_rest"},
{'params': [self._opacity], 'lr': training_args.opacity_lr, "name": "opacity"},
{'params': [self._scaling], 'lr': training_args.scaling_lr, "name": "scaling"},
{'params': [self._rotation], 'lr': training_args.rotation_lr, "name": "rotation"}
]
# 创建optimizer
self.optimizer = torch.optim.Adam(l, lr=0.0, eps=1e-15)
# 创建对xyz参数进行学习率调整的scheduler
self.xyz_scheduler_args = get_expon_lr_func(lr_init=training_args.position_lr_init*self.spatial_lr_scale,
lr_final=training_args.position_lr_final*self.spatial_lr_scale,
lr_delay_mult=training_args.position_lr_delay_mult,
max_steps=training_args.position_lr_max_steps)
scene/dataset_readers.py
class SceneInfo(NamedTuple):
point_cloud: BasicPointCloud
train_cameras: list
test_cameras: list
nerf_normalization: dict
ply_path: str
#计算这些相机的平均中心的负值和所在球半径大小
def getNerfppNorm(cam_info):
#计算这些相机的平均中心,以及到平均中心的最远距离,返回平均中心和最远距离
def get_center_and_diag(cam_centers):
cam_centers = np.hstack(cam_centers)
avg_cam_center = np.mean(cam_centers, axis=1, keepdims=True)
center = avg_cam_center
dist = np.linalg.norm(cam_centers - center, axis=0, keepdims=True)
diagonal = np.max(dist)
return center.flatten(), diagonal
cam_centers = []
#计算相机在世界坐标系的位置,并放入cam_centers列表中
for cam in cam_info:
W2C = getWorld2View2(cam.R, cam.T)
C2W = np.linalg.inv(W2C)
cam_centers.append(C2W[:3, 3:4])
center, diagonal = get_center_and_diag(cam_centers)
radius = diagonal * 1.1
translate = -center
return {"translate": translate, "radius": radius}
#读取相机信息,参数:相机外参,内参,图片目录,输出到列表中,列表每一项是CameraInfo类
def readColmapCameras(cam_extrinsics, cam_intrinsics, images_folder):
cam_infos = []
for idx, key in enumerate(cam_extrinsics):
sys.stdout.write('\r')
# the exact output you're looking for:
sys.stdout.write("Reading camera {}/{}".format(idx+1, len(cam_extrinsics)))
sys.stdout.flush()
extr = cam_extrinsics[key]
intr = cam_intrinsics[extr.camera_id]
height = intr.height
width = intr.width
uid = intr.id
R = np.transpose(qvec2rotmat(extr.qvec))
T = np.array(extr.tvec)
if intr.model=="SIMPLE_PINHOLE":
focal_length_x = intr.params[0]
FovY = focal2fov(focal_length_x, height)
FovX = focal2fov(focal_length_x, width)
elif intr.model=="PINHOLE":
focal_length_x = intr.params[0]
focal_length_y = intr.params[1]
FovY = focal2fov(focal_length_y, height)
FovX = focal2fov(focal_length_x, width)
else:
assert False, "Colmap camera model not handled: only undistorted datasets (PINHOLE or SIMPLE_PINHOLE cameras) supported!"
image_path = os.path.join(images_folder, os.path.basename(extr.name))
image_name = os.path.basename(image_path).split(".")[0]
image = Image.open(image_path)
cam_info = CameraInfo(uid=uid, R=R, T=T, FovY=FovY, FovX=FovX, image=image,
image_path=image_path, image_name=image_name, width=width, height=height)
cam_infos.append(cam_info)
sys.stdout.write('\n')
return cam_infos
#读取points3D.ply文件,读取位置、颜色和法线,并组合成BasicPointCloud类返回
def fetchPly(path):
plydata = PlyData.read(path)
vertices = plydata['vertex']
positions = np.vstack([vertices['x'], vertices['y'], vertices['z']]).T
colors = np.vstack([vertices['red'], vertices['green'], vertices['blue']]).T / 255.0
normals = np.vstack([vertices['nx'], vertices['ny'], vertices['nz']]).T
return BasicPointCloud(points=positions, colors=colors, normals=normals)
def storePly(path, xyz, rgb):
# Define the dtype for the structured array
dtype = [('x', 'f4'), ('y', 'f4'), ('z', 'f4'),
('nx', 'f4'), ('ny', 'f4'), ('nz', 'f4'),
('red', 'u1'), ('green', 'u1'), ('blue', 'u1')]
normals = np.zeros_like(xyz)
elements = np.empty(xyz.shape[0], dtype=dtype)
attributes = np.concatenate((xyz, normals, rgb), axis=1)
elements[:] = list(map(tuple, attributes))
# Create the PlyData object and write to file
vertex_element = PlyElement.describe(elements, 'vertex')
ply_data = PlyData([vertex_element])
ply_data.write(path)
def readColmapSceneInfo(path, images, eval, llffhold=8):
try:
cameras_extrinsic_file = os.path.join(path, "sparse/0", "images.bin")
cameras_intrinsic_file = os.path.join(path, "sparse/0", "cameras.bin")
cam_extrinsics = read_extrinsics_binary(cameras_extrinsic_file)
cam_intrinsics = read_intrinsics_binary(cameras_intrinsic_file)
except:
cameras_extrinsic_file = os.path.join(path, "sparse/0", "images.txt")
cameras_intrinsic_file = os.path.join(path, "sparse/0", "cameras.txt")
cam_extrinsics = read_extrinsics_text(cameras_extrinsic_file)
cam_intrinsics = read_intrinsics_text(cameras_intrinsic_file)
reading_dir = "images" if images == None else images
cam_infos_unsorted = readColmapCameras(cam_extrinsics=cam_extrinsics, cam_intrinsics=cam_intrinsics, images_folder=os.path.join(path, reading_dir))
cam_infos = sorted(cam_infos_unsorted.copy(), key = lambda x : x.image_name)
#需要评估的话,训练集和测试集分一下,否则全部都拿来训练
if eval:
train_cam_infos = [c for idx, c in enumerate(cam_infos) if idx % llffhold != 0]
test_cam_infos = [c for idx, c in enumerate(cam_infos) if idx % llffhold == 0]
else:
train_cam_infos = cam_infos
test_cam_infos = []
nerf_normalization = getNerfppNorm(train_cam_infos)
ply_path = os.path.join(path, "sparse/0/points3D.ply")
bin_path = os.path.join(path, "sparse/0/points3D.bin")
txt_path = os.path.join(path, "sparse/0/points3D.txt")
if not os.path.exists(ply_path):
print("Converting point3d.bin to .ply, will happen only the first time you open the scene.")
try:
xyz, rgb, _ = read_points3D_binary(bin_path)
except:
xyz, rgb, _ = read_points3D_text(txt_path)
storePly(ply_path, xyz, rgb)
try:
#读取ply文件
pcd = fetchPly(ply_path)
except:
pcd = None
#把这些组合成SceneInfo类返回
scene_info = SceneInfo(point_cloud=pcd,
train_cameras=train_cam_infos,
test_cameras=test_cam_infos,
nerf_normalization=nerf_normalization,
ply_path=ply_path)
return scene_info
def readCamerasFromTransforms(path, transformsfile, white_background, extension=".png"):
cam_infos = []
with open(os.path.join(path, transformsfile)) as json_file:
contents = json.load(json_file)
fovx = contents["camera_angle_x"]
frames = contents["frames"]
for idx, frame in enumerate(frames):
cam_name = os.path.join(path, frame["file_path"] + extension)
# NeRF 'transform_matrix' is a camera-to-world transform
c2w = np.array(frame["transform_matrix"])
# change from OpenGL/Blender camera axes (Y up, Z back) to COLMAP (Y down, Z forward)
c2w[:3, 1:3] *= -1
# get the world-to-camera transform and set R, T
w2c = np.linalg.inv(c2w)
R = np.transpose(w2c[:3,:3]) # R is stored transposed due to 'glm' in CUDA code
T = w2c[:3, 3]
image_path = os.path.join(path, cam_name)
image_name = Path(cam_name).stem
image = Image.open(image_path)
im_data = np.array(image.convert("RGBA"))
bg = np.array([1,1,1]) if white_background else np.array([0, 0, 0])
norm_data = im_data / 255.0
arr = norm_data[:,:,:3] * norm_data[:, :, 3:4] + bg * (1 - norm_data[:, :, 3:4])
image = Image.fromarray(np.array(arr*255.0, dtype=np.byte), "RGB")
fovy = focal2fov(fov2focal(fovx, image.size[0]), image.size[1])
FovY = fovy
FovX = fovx
cam_infos.append(CameraInfo(uid=idx, R=R, T=T, FovY=FovY, FovX=FovX, image=image,
image_path=image_path, image_name=image_name, width=image.size[0], height=image.size[1]))
return cam_infos
def readNerfSyntheticInfo(path, white_background, eval, extension=".png"):
print("Reading Training Transforms")
train_cam_infos = readCamerasFromTransforms(path, "transforms_train.json", white_background, extension)
print("Reading Test Transforms")
test_cam_infos = readCamerasFromTransforms(path, "transforms_test.json", white_background, extension)
if not eval:
train_cam_infos.extend(test_cam_infos)
test_cam_infos = []
nerf_normalization = getNerfppNorm(train_cam_infos)
ply_path = os.path.join(path, "points3d.ply")
if not os.path.exists(ply_path):
# Since this data set has no colmap data, we start with random points
num_pts = 100_000
print(f"Generating random point cloud ({num_pts})...")
# We create random points inside the bounds of the synthetic Blender scenes
xyz = np.random.random((num_pts, 3)) * 2.6 - 1.3
shs = np.random.random((num_pts, 3)) / 255.0
pcd = BasicPointCloud(points=xyz, colors=SH2RGB(shs), normals=np.zeros((num_pts, 3)))
storePly(ply_path, xyz, SH2RGB(shs) * 255)
try:
pcd = fetchPly(ply_path)
except:
pcd = None
scene_info = SceneInfo(point_cloud=pcd,
train_cameras=train_cam_infos,
test_cameras=test_cam_infos,
nerf_normalization=nerf_normalization,
ply_path=ply_path)
return scene_info
#字典,存储对应的回调函数
sceneLoadTypeCallbacks = {
"Colmap": readColmapSceneInfo,
"Blender" : readNerfSyntheticInfo
}