Load a class from a module path. Args: module_path: Python module path (e.g., 'datasets.seven_scenes'). class_name: C

Load intrinsics, returns dict mapping frame index (int) to [fx,fy,cx,cy]. Args: loader: BSSLoader instance Returns:

Read poses.txt and return all C2W matrices in OpenCV convention. poses.txt contains 4×4 matrices stored 4 lines each (row-major), in

Load and cache all C2W poses from poses_c2w.txt. Each line contains 16 space-separated floats (4x4 matrix, row-major). Returns shape

Lazy-load and cache per-scene RGB file list and GT trajectory.

One-line dump of aggregator KV cache occupancy + aux counters.

Normalize a single raw score map to uint8 [0, 255].

Parse COLMAP sparse reconstruction into poses + intrinsics.

Parse COLMAP text-format cameras.txt + images.txt. Returns dict with: 'poses': {image_name: 4x4 C2W float32} 'intrin

Ensure the sky mask cache directory exists and write the version stamp.

Preprocess a single BGR image for skyseg ONNX model. Returns (3, H, W) float32.

Regenerate camera visualizations with current settings.

Regenerate all point clouds with current settings.

Render to image, apply EDL if depth provided. When EDL is active, uses point-cloud-only depth (edl_depth) for shading and a second de

Fallback rendering when screenshot capture fails.

Apply overlays then render.

Reset camera to view the first frame.

Rollback the most recent frame from all caches. Undoes append_frame on FlashInfer manager (all blocks), trims the camera head SDPA ca

Rollback the most recent frame from all caches. Undoes append_frame on FlashInfer manager (all blocks), trims the camera head SDPA ca

Read-modify-write: read existing json, update method key, write back.

Compute relative pose errors between frame pairs (W2C format).

Smoothly transition camera to target pose using look_at based control. Args: client: Viser client handle. target_posi

Convert module name to expected class name. Converts snake_case to PascalCase and appends 'Dataset' or 'Method'. Examples:

Enable/disable the Clear Cache button based on whether a cache exists.

Drive `clean_kv_cache → Phase 1 → N streaming forwards` `passes` times. Warmup inputs are sliced from the already-preprocessed ``images`` tensor,

Activate pose parameters with specified activation functions. Args: pred_pose_enc: Tensor containing encoded pose parameters [transl

Add camera visualization for a frame.

Adjust camera intrinsics for a ResizeTransform. Applies: 1. Scale: fx *= scale_x, fy *= scale_y, cx *= scale_x, cy *= scale_y 2.

Returns (vis_xyz, vis_rgb, extra_geoms).

Aligns the 3D scene based on the extrinsics of the first camera. Args: scene_3d: The 3D scene to be aligned extrinsics_matri

Apply sky segmentation to confidence scores. Args: conf: Confidence scores with shape (S, H, W) image_folder: Path to the fo

Build a CameraPath from CameraConfig segments. Args: cam_config: CameraConfig with fov, transition, segments scene: Scene with c2

Build overlay instances + serializable specs from config. Reads ``cfg.overlay`` and the rendered scene, returning ``(overlays, specs)`` suita

Remove all directory contents unconditionally (force mode).

支持任意 batch 维度的 SE3 逆运算 se3: (..., 4, 4) 或 (..., 3, 4)

Colorize points by distance error using a colormap. Maps nearest-neighbor distances to colors, capped at 3*tau. Ported from the official TAT

Scene world-up from median of camera -Y axes.

Custom interpolate to avoid INT_MAX issues in nn.functional.interpolate.

Convert a float32 depth map to a BGR pseudo-colour image (COLORMAP_TURBO). Invalid pixels (depth == 0) are rendered as black. Args:

Release octree references.

Encode original RGB frames (S,H,W,3 uint8) to MP4. If *width* or *height* is given, frames are resized before encoding. Providing only one di

Evaluate a single frame. Args: pred_depth: Predicted depth map gt_depth: Ground truth depth map Returns:

Run the eviction that was skipped while ``_defer_eviction`` was True.

Export predictions to GLB.

Find a folder that contains images inside a scene path.

Derive pinhole intrinsics (fx, fy, cx, cy) from vertical FOV.

Load config from YAML file.

Read-only snapshot of cache occupancy for one block. Useful for debugging keyframe / sliding-window behavior. Returns:

Get evaluation config merged for a specific dataset and method. Merge order (later overrides earlier): 1. base.yaml evaluation (g

Get configuration for a specific method. Returns: Dictionary with: - 'method_class': Method module name (e.g., 'da3')

Returns the OpenGL conversion matrix (flips Y and Z axes).

Get selected method names from the 'methods' list in base config. Returns an empty list if 'methods' is absent or empty. Raises KeyEr

Check whether a frame-level data key is present.

Integrates a camera mesh into the 3D scene. Args: scene: The 3D scene to add the camera model transform: Transformation matr

Load images from folder or video and preprocess into a tensor.

Load and preprocess images from an explicit path list.

Load camera intrinsics. Returns: np.ndarray of shape (N, 4) with [fx, fy, cx, cy] per row. NaN rows where intrinsics

Load a method class by name. Searches for the class in the methods/ directory. Args: name: Method name (e.g., 'passthrou

Load GCTStream model from checkpoint.

Load cached sky masks or generate them with the ONNX model. Args: image_folder: Folder containing input images. image_paths:

Load predictions from NPZ file or directory of per-frame npz files. Per-frame NPZ directories are loaded in parallel using a thread pool — np

Generate follow-camera path as keyframes. Args: keyframe_interval: Generate a keyframe every N frames. 1 = every frame (lossl

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