Run aggregator to get multi-scale features. Args: images: Input images [B, S, 3, H, W] num_frame_for_scale:

Run aggregator to get multi-scale features. Args: images: Input images [B, S, 3, H, W] num_frame_for_scale:

Run aggregator to get multi-scale features. Args: images: Input images [B, S, 3, H, W] num_frame_for_scale:

Build streaming aggregator with KV cache support (FlashInfer backend). Returns: AggregatorStream module

Build streaming aggregator with KV cache support (FlashInfer backend). Returns: AggregatorStream module

Build streaming aggregator with KV cache support (FlashInfer backend). Returns: AggregatorStream module

Build frame and global blocks for streaming causal mode.

Build causal camera head for streaming inference. Returns: CameraCausalHead module or None

Build causal camera head for streaming inference. Returns: CameraCausalHead module or None

Build causal camera head for streaming inference. Returns: CameraCausalHead module or None

Setup client connection callbacks.

Handle numpy types in JSON serialization.

Load one per-frame npz, keeping only keys the renderer needs.

Prepare camera, register, and scale tokens. Args: B: Batch size S_local: Local sequence length S

Process causal global attention via FlashInfer streaming path. Args: tokens: Input tokens B: Batch size

Rebuild the 3D scene from currently selected data.

Save confidence map as EXR, resized to GT resolution.

Save custom frame data using context's __save_{key}_file__ method.

Save depth map, resized to GT resolution with nearest-neighbor.

Save mask image, resized to GT resolution with nearest-neighbor.

Save per-frame world-coordinate point grid as EXR, resized to GT resolution.

Save RGB image, resized to GT resolution.

Setup camera, register, and scale tokens for causal mode.

Spherical linear interpolation between quaternions.

Consumer worker: reads visible points from SharedMemory, renders.

Compute accuracy: mean distance from source to target. Measures how close the source points are to the target point cloud. Lower values indic

Return (xyz, rgb, frame_indices) up to frame_idx. No culling.

Adjust camera intrinsics for this resize transform. Note: get_transform() must have been called first to initialize the cached transf

compute distance matrix between all pairs of poses

Calculate the Area Under the Curve (AUC) for the given error arrays using NumPy. Args: r_error: numpy array representing R error val

Normalize predicted tracks based on camera intrinsics. Args: intrinsics (torch.Tensor): The camera intrinsics tensor of shape [batch_size

Actual up vector (orthogonal to forward and right).

Clean KV cache (call this when starting a new sequence).

Modify camera intrinsics to follow a different convention. Coordinates of the center of the top-left pixels are by default: - (0.5, 0.5)

Turn a grayscale image into a color image (PyTorch tensor version). Args: x: Grayscale image tensor [H, W] or [B, H, W] cmap

Compute completeness: mean distance from target to source. Measures how well the source points cover the target point cloud. Lower values ind

Computes the faces for the camera mesh.

Compute depth range across multiple depth maps. Uses 1st and 99th percentiles to exclude outliers. Args: depths: List of HxW depth m

compute flow magnitude between all pairs of frames

compute flow magnitude between all pairs of frames

Convert Point3D extrinsic matrices to OpenCV convention. Args: Rot: 3D rotation matrix in Point3D format Trans: 3D translati

Create onnxruntime.InferenceSession; raises ImportError if not installed. Args: model_path: Path to ONNX model file. Returns:

Convert depth map to point cloud using camera intrinsics. This function backprojects a depth map to 3D space using pinhole camera model. Poin

Return a display name for GUI.

Whether resizing is active (mode != 'none').

Evaluate depth predictions against ground truth. Args: gt_loader: BSSLoader for the ground truth directory pred_loa

Run point cloud evaluation via dataset.evaluate_pointcloud(). Args: gt_loader: BSSLoader for ground truth directory

Compute pose AUC metrics for a scene. Args: gt_loader: BSSLoader for ground truth directory pred_loader: BSSLoader

Evaluate trajectory using evo library (ATE and RPE). Args: gt_loader: BSSLoader for ground truth directory pred_loa

Compute comprehensive point cloud evaluation metrics. Convenience function that computes all common metrics in one call. Args: sourc