3DGS本身支持对 Blender 数据集的训练，其主要数据格式为：
<location>
|---images
| |---<image 0>
| |---<image 1>
| |---...
|---transformers_train.json
|---point3d.ply
数据准备#
通常我们自己采集的数据集来自激光扫描装置，提供了 las 格式的点云文件、json 格式的相机位姿以及图片。
首先处理 las 格式的点云文件，需要将其转化为二进制的 ply 文件。
然后查看 transformers_train.json 文件，主要是针对性修改 3dgs 代码文件下 /sence/dataset_readers.py 这个文件。
def readNerfSyntheticInfo(path, white_background, eval, extension=".jpg"):
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 = []
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
注释掉了 Test Transforms 相关的部分，不会进行 eval，因此不需要 test。修改了 extension，主要是看自己的图像输入格式。
def readCamerasFromTransforms(path, transformsfile, white_background, extension=".jpg"):
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,"images", frame["file_path"])
# 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
这里通常根据自己的 json 文件内容更改，几个主要需要注意的点是：
-
fovx = contents["camera_angle_x"] -
cam_name = os.path.join(path,"images", frame["file_path"])
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([0,0,0]).T
return BasicPointCloud(points=positions, colors=colors, normals=normals)
这部分是 ply 文件相关的，通常自采集数据集不会有点云法向量 normals，因此我们根据其 colmap 部分的代码，也将其设置为 0。
一般这样改完，代码就能跑通了，激光点云结合 GPS 高精度相机位姿，能够解决 colmap 无法获取相机位姿的问题。