Aggregate a list of metric dicts, computing mean/median/std. Non-finite values (NaN / +/-Inf) are treated as bad cases and excluded. Typical sour

Align point clouds of two overlapping chunks: point_map2 → point_map1 coordinate frame. Args: point_map1: (B, H, W, 3) world-coordinate p

Align estimated trajectory to reference using Umeyama Sim(3). Default no RANSAC; if ransac=True, use RANSAC (max iterations default 10).

Align estimated trajectory to reference using Umeyama Sim(3). Default no RANSAC; if ransac=True, use RANSAC (max iterations default 10).

Apply exponential moving average smoothing over time.

Apply specified activation function to parameter tensor Args: tensor: Tensor containing parameter values

Apply SIM(3) transform to cam2world 4x4 pose matrices. Args: c2w_44: (N, 4, 4) cam2world homogeneous matrices s: scale R:

Apply basic activation function to pose parameters. Args: pose_enc: Tensor containing encoded pose parameters act_type: Acti

Apply basic activation function to pose parameters. Args: pose_enc: Tensor containing encoded pose parameters act_type: Acti

Apply basic activation function to pose parameters. Args: pose_enc: Tensor containing encoded pose parameters act_type: Acti

Apply basic activation function to pose parameters. Args: pose_enc: Tensor containing encoded pose parameters act_type: Acti

Apply basic activation function to pose parameters. Args: pose_enc: Tensor containing encoded pose parameters act_type: Acti

Apply basic activation function to pose parameters. Args: pose_enc: Tensor containing encoded pose parameters act_type: Acti

Apply basic activation function to pose parameters. Args: pose_enc: Tensor containing encoded pose parameters act_type: Acti

Apply basic activation function to pose parameters. Args: pose_enc: Tensor containing encoded pose parameters act_type: Acti

Apply basic activation function to pose parameters. Args: pose_enc: Tensor containing encoded pose parameters act_type: Acti

Return page IDs in strict order: scale → window → special. Placing special pages last means only the final page may be partially

Method: 1) express all 3d points in each camera coordinate frame 2) if they're in front of a depthmap --> then lower their confidence

Valid token count in the last page of the visible sequence. - No special pages → last page is a patch page.

Compute non-sky mask from sky prediction. Args: sky_prediction: Sky prediction tensor threshold: Threshold for sky classific

Compute non-sky mask from sky prediction. Args: sky_prediction: Sky prediction tensor threshold: Threshold for sky classific

Convert general model state dict to match current model architecture. Args: state_dict: Original state dictionary Returns:

Convert ray directions, depth along ray, pose translation, and unit quaternions (representing pose rotation) to a pointmap in world frame.

Args: - depthmap (BxHxW array): - pseudo_focal: [B,H,W] ; [B,2,H,W] or [B,1,H,W] Returns: pointmap of absolute coordi

Reprojection method, for when the absolute depth is known: 1) estimate the camera focal using a robust estimator 2) reproject points

Evaluate a single scene. Args: eval_metrics: list of metric categories to evaluate. Allowed values: - "depth": depth metrics

Convert camera extrinsics and intrinsics to a compact pose encoding.

Forward pass of AMB3R 1) front-end processing 2) back-end processing

Generate a coordinate frame given perpendicular, unit-length Y and Z vectors.

Generate a coordinate frame given perpendicular, unit-length Y and Z vectors.

Get a registered model adapter by name.

Get stored keyframe pose encoding.

Convert camera intrinsics to a raymap (ray origins + directions) in camera frame. Note: Currently only supports pinhole camera model. Ar

Append a '1' along the final dimension of the tensor (i.e. convert xyz->xyz1)

Convert batched points (xyz) to (xyz1).

Convert batched points (xyz) to (xyz1).

Integrates a fake camera mesh into the 3D scene. Args: scene (trimesh.Scene): The 3D scene to add the camera model. transfor

Apply inverse log transform: sign(y) * (exp(|y|) - 1) Args: y: Input tensor Returns: Transformed tensor

Apply inverse log transform: sign(y) * (exp(|y|) - 1) Args: y: Input tensor Returns: Transformed tensor

Apply inverse log transform: sign(y) * (exp(|y|) - 1) Args: y: Input tensor Returns: Transformed tensor

Apply inverse log transform: sign(y) * (exp(|y|) - 1) Args: y: Input tensor Returns: Transformed tensor

Apply inverse log transform: sign(y) * (exp(|y|) - 1) Args: y: Input tensor Returns: Transformed tensor

Apply inverse log transform: sign(y) * (exp(|y|) - 1) Args: y: Input tensor Returns: Transformed tensor

Apply inverse log transform: sign(y) * (exp(|y|) - 1) Args: y: Input tensor Returns: Transformed tensor

Apply inverse log transform: sign(y) * (exp(|y|) - 1) Args: y: Input tensor Returns: Transformed tensor

Apply inverse log transform: sign(y) * (exp(|y|) - 1) Args: y: Input tensor Returns: Transformed tensor

Apply inverse log transform: sign(y) * (exp(|y|) - 1) Args: y: Input tensor Returns: Transformed tensor

Load a configuration. Will resolve inheritance. Supports both file paths and module paths (e.g., depth_anything_3.configs.giant).

Load a configuration. Will resolve inheritance. Supports both file paths and module paths (e.g., depth_anything_3.configs.giant).

Add keyframes using fixed-interval or novelty-gated selection. `warmup_confidence` is the current segment's warmup-window mean post-softplus

Called after fallback procedure finishes. Sets cooldown.

Executes a given function in parallel using threads or processes. When using threads, the parallelism is achieved during IO blocking.

Convert a pose encoding back to camera extrinsics and intrinsics. This function performs the inverse operation of extri_intri_to_pose_encoding,

Convert a pose encoding back to camera extrinsics and intrinsics.

Prepare Gaussian splats from model predictions and input data. Args: views: Dictionary containing view data (cam

Convert a quaternion into a 3x3 rotation matrix. Args: - quat: 4 or Bx4 torch tensor (unit quaternions and notation is (x, y, z, w))

Reconstruct absolute camera sequence from relative predictions.

extract confidence from prediction head output

Clear all state.

Build flat residual vector for LM: translation, rotation, and prior terms.