Given an equirectangular RGB-D image, back-project each pixel to a 3D point and compute the corresponding 3D Gaussian covariance so that the

Converts panoramic RGBD data (image, distance, rays) into an Open3D mesh. Args: rgb: Input RGB image tensor (M, H, W, 3), uint8 or f

Given an equirectangular RGB-D image, back-project each pixel to a 3D point and compute the corresponding 3D Gaussian covariance so that the

Converts panoramic RGBD data (image, distance, rays) into an Open3D mesh. Args: image: Input RGB image tensor (H, W, 3), uint8 or fl

将权重矩阵转换为 one-hot 形式。对于每个像素，选择最大权重对应的索引位置为 1，其他位置为 0。 :param weights: 形状为 [B, 4, H, W] 的权重矩阵 :return: 转换后的 one-hot 矩阵，形状为 [B, 4, H, W]

将权重矩阵转换为 one-hot 形式。对于每个像素，选择最大权重对应的索引位置为 1，其他位置为 0。 :param weights: 形状为 [B, 4, H, W] 的权重矩阵 :return: 转换后的 one-hot 矩阵，形状为 [B, 4, H, W]

将输入的权重矩阵中每个位置上最小的两个值置0，然后重新归一化。 :param weights: 输入的权重矩阵，形状为 [B, 4, H, W] :return: 处理后的权重矩阵，形状为 [B, 4, H, W]

Rotation from global (world) to camera coordinates - camera coordinate has the z-axis as forward and x- and y-axis for image. - when we rotat

Rotation from global (world) to camera coordinates

Create coordinate grid with height and width NOTE: primarly used for equi2pers `z-axis` scale is `1` params: - height (int) - w

Create coordinate grid with height and width `z-axis` scale is `1` params: - height (int) - width (int) - batch (Optional[int])

Create intrinsic matrix params: - height, width (int) - fov_x (float): make sure it's in degrees - skew (float): 0.0 - dtype (np.

Create intrinsic matrix params: - height, width (int) - fov_x (float): make sure it's in degrees - skew (float): 0.0 - dtype (tor

Create coordinate grid with height and width NOTE: primarly used for equi2equi params: - height (int) - width (int) - batch (Opt

Create coordinate grid with height and width NOTE: primarly used for equi2equi params: - height (int) - width (int) - batch (Opt

Create Rotation Matrix params: - roll, pitch, yaw (float): in radians - z_down (bool): flips pitch and yaw directions - dtype (np.dty

Create Rotation Matrix params: - roll, pitch, yaw (float): in radians - z_down (bool): flips pitch and yaw directions - dtype (torch.

xyz coordinates of the faces of the cube

xyz coordinates of the faces of the cube

Crops an image to simulate reduced horizontal and vertical field of views (FOV) based on angles. Parameters: - image: PIL.Image.Imag

params: - cubemap - cube_format (str): ("dice", "horizon", "dict", "list") - height, width (int): output size - mode (str): "bili

Function that converts angles from degrees to radians

Args: normalized_depth: [B, ...], torch.FloatTensor, [-1.0, 1.0] min_value: [B,], torch.FloatTensor, [0.0, +inf]

将深度图从直线距离转换为z距离。 :param depth_map: 输入深度图，形状为 (..., H, W, 1) 的 np.ndarray :param fov_x: 相机的水平视场角（单位：度）。 :param fov_y: 相机的垂直视场角（单位：度）。

params: - equi (np.ndarray, torch.Tensor) - rot (dict, list[dict]): {"roll", "pitch", "yaw"} - w_face (int): cube face width - cu

params: - src - rots - mode (str): interpolation mode, defaults to "bilinear" - z_down (bool) - height, width (optional int):

params: - equi - rots - height, width (int): perspective size - fov_x (float): perspective image fov of x-axis - z_down (bool

Generate triangle indices for a closed cube mesh from a CubeMap of shape [6, H, W]. Returns: [N, 3] triangle index tensor

Args: L: cube face length, int Returns: mask: np.ndarray, (3L, 4L), uint8

Inference function for the DreamCube model. Args: cube_rgbs (torch.Tensor): Input RGB images of shape (B, M, C, H, W). cube_d

计算掩码为True的元素, 对应的最大值和最小值。 参数: input_tensor (torch.Tensor): 形状为 [B, ...] 的输入张量。 mask (torch.Tensor): 形状为 [B, ...] 的布尔

Nearest Neighbor Interpolation Merit of using this nearest instead is that the grid doesn't need to be a cuda tensor. Although it is a little

Nearest Neightbor Sampling

Args: image: np.ndarray, [H, W, C]

Postprocess the generated images. Args: depths: (bs, m, h, w, 1), np.float32 Returns: dict: postprocessed depths in diffe

Postprocess the generated images. Args: images: (bs, m, h, w, 3), np.uint8 Returns: dict: postprocessed images in differe

r"""Resample padding function. Replaces padded regions of cube faces with content from neighboring faces. Args: x: Inpu

对 [B, C, K, K, H, W] 形状的 patches 进行 2D 旋转 :param patches: Tensor, shape [B, C, K, K, H, W] :param theta: Tensor, shape [H, W], 角度（弧度制）

对 [B, C, K, K, H, W] 形状的 patches 进行 2D 旋转 :param patches: Tensor, shape [B, C, K, K, H, W] :param theta: Tensor, shape [H, W], 角度（弧度制）

patches: [B, C, K, K, H, W] theta: int

参数: weight (torch.Tensor): 输入的卷积核权重，形状为 [C_in, C_out, K, K] 返回: torch.Tensor: 旋转后的卷积核权重，形状相同。

Visualizes depth maps, such as predictions of the `MarigoldDepthPipeline`. Args: depth (`Union[PIL.Image.Image, np.ndarr

将z距离图转换为直线距离。 :param z_distance: 输入 z 距离图，形状为 (..., H, W, 1) 的 np.ndarray :param fov_x: 相机的水平视场角（单位：度）。 :param fov_y: 相机的垂直视场角（单位：度）