Generate a 3D point cloud and camera wireframes and export them as a ``.glb`` file. The function builds a point cloud from the predicted depth ma

Convert the extrinsics to a representation with 5 degrees of freedom: 1. Distance from pivot point in the "X" (look cross pivot axis) direction.

Batch process RGB arrays (N, 3) factor: 0.0 (grayscale) to 1.0 (original color)

Retrieves or interpolates temporal position embeddings.

Linearly interpolate short missing segments in trajectories.

Batch version of weighted least squares Homography src_pts_batch: (B, K, 2) dst_pts_batch: (B, K, 2) confident_weight_batch: (B, K)

Convert focal length (in pixels) to field-of-view (radians). Args: focal: focal length along x or y axis Returns: fov:

Unified feature and geometry extraction interface. - for_flow: If False, only extract and return geometric outputs (Pure Point mode)

embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D)

Calculates necessary padding for time dimension to fit window size S.

this will perform the index select, cat the tensors, and provide the attn_bias from cache

this will perform the index select, cat the tensors, and provide the attn_bias from cache

this will perform the index select, cat the tensors, and provide the attn_bias from cache

Get scene data including images, camera parameters, and auxiliary info. Args: scene: Scene identifier Returns:

Collect per-view image paths, intrinsics/extrinsics for a scene. Only uses iPhone images (not DSLR). Args: scen

Collect per-view image paths, intrinsics/extrinsics for a scene. Args: scene: Scene identifier (e.g., "chess")

Collect per-view image paths, intrinsics/extrinsics for a scene. Args: scene: Scene path (e.g., "xxx/yyy/zzz")

Collect per-view image paths, intrinsics/extrinsics for a scene. Args: scene: Scene identifier (e.g., "courtyard")

Wrapper to perform PCA on features and normalize.

Constructs transform from Reference to Memory (Voxel) coordinates.

Defines the superset of arguments required for both tasks.

Get backend status information.

Get view data by index, handling bounds. Args: processed_data: Processed data dictionary view_index: Index o

Combines a sequence of images into a GIF (left) and Tiled view (right).

Convert batched points (xyz) to (xyz1).

Import a python item. Example: import_item("path.to.file", "MyClass") -> MyClass

Perform standard inference on a batch of video frames and output metric 3D geometry. Compared to the evaluation routine, this functi

Run inference according to requested modes. - Unposed export if 'pose' or 'recon_unposed' is in modes - Posed export if 'rec

Pairwise inference function designed for consecutive-frame processing. In contrast to the standard `infer` interface (which focuses

Interpolate extrinsics by rotating around their "focus point," which is the least-squares intersection between the look vectors of the initial and

Launch the application. Args: host: Host address to bind to port: Port number to bind to **kwarg

Loads 16-bit PNG depth from Kubric dataset and converts to meters.

Load a scene from examples directory. Args: scene_name: Name of the scene to load examples_dir: Path to exam

Load all metrics JSON files from a directory. Args: metric_dir: Path to directory containing metric JSON files Returns:

Load model if not already loaded.

Load track4world checkpoint weights with key remapping for nested DA3. The original track4world_da3.pth stores DA3 backbone weights

This function generates a 1D positional embedding from a given grid using sine and cosine functions. # noqa Args: - embed_dim: The embed

Convert rotations given as rotation matrices to quaternions. Args: matrix: Rotation matrices as tensor of shape (..., 3, 3). Re

Convert rotations given as rotation matrices to quaternions. Args: matrix: Rotation matrices as tensor of shape (..., 3, 3). Re

Returns 2D meshgrid (Y, X).

Navigate measure view. Args: processed_data: Processed data dictionary current_selector: Current selector va

UV with left-top corner as (-width / diagonal, -height / diagonal) and right-bottom corner as (width / diagonal, height / diagonal)

Generate normalized UV coordinates on the camera view plane. The coordinates are in a normalized range based on the image diagonal: - To

Converts a 1-channel tensor to a 3-channel heatmap (Inferno) or grayscale. Input: (B, 1, H, W) or (B, H, W) Output: (B, 3, H, W) uint8 te

Convert 2D position grid (HxWx2) to sinusoidal embeddings (HxWxC) Args: pos_grid: Tensor of shape (H, W, 2) containing 2D coordinate

Converts [0, 255] uint8/float to [-0.5, 0.5] float.

Print evaluation metrics in a beautiful tabular format. Args: metrics: Dictionary mapping "dataset_mode" to metric resul

Filter unstable 3D trajectories based on visibility continuity, motion jumps, and smoothness constraints, and optionally prepare them for smo

Quaternion Order: XYZW or say ijkr, scalar-last Convert rotations given as quaternions to rotation matrices. Args: quaternions:

Recover the depth map and FoV from a point map with unknown z shift and focal. Note that it assumes: - the optical center is at the cent

Removes outliers using Open3D (CPU implementation).

Reorder views to place the selected reference view first. Args: x: Input tensor of shape (B, S, N, C) b_idx: Reference v

Generic pipeline for running inference and evaluation. Args: args: Parsed arguments. task_config (dict): Configuration s

Computes the inverse of a batch of SE(3) matrices.

Set sky regions to maximum depth and high confidence. Args: depth: Depth tensor depth_conf: Depth confidence tensor

Smooth trajectories in the temporal dimension using Gaussian filtering for extremely high smoothness. Args: trajs: (T, N, 3) Tra

Start the backend server.

Summarizes a list of images as a GIF/Tile.

frame: (H,W,D) -> (H,W,3)

Sim(3) transform point cloud points: (N, 3) rot: (3, 3) trans: (3,) or (1, 3) scale: float inverse: Whether to do inverse tra

for np < 2

Update depth view for a specific view index. Args: processed_data: Processed data dictionary view_index: Ind

Update measure view for a specific view index. Args: processed_data: Processed data dictionary view_index: I

Updates the trajectory lines based on current timestep.

Upsample low-resolution flow using learned convex mask. Standard RAFT upsampling.

Compute weighted mean of a numpy array. Args: x: input array w: optional weights of the same shape as x axis: axis o

Wraps angle to [-pi, pi].

Initializes the position generator with an empty cache.

Initialize DepthAnything3Net with given yaml-initialized configuration.

Initializes the position generator with an empty cache.

Draws camera frustums to visualize their position and orientation. This function renders each camera as a wireframe pyramid, originating from

Add processed images to prediction for visualization.

Align depth map to input extrinsics

Apply depth alignment using least squares scaling.

Apply metric scaling to the metric depth output.