Sanitize scene name for use in file system (replaces '/' with '_'). Example: >>> BSSManager.resolve_scene_safe("chess/seq-01")

Save evaluation results to eval/{eval_type}.json.

Setup unified logging system with file and console handlers. Args: log_dir: Directory for log files name: Logger name (e.g., 'pre

Downsample a point cloud using a voxel grid filter. Args: points: (N, 3) or (N, 6) xyz or xyzrgb point cloud. voxel_size: Voxel s

W2C (4x4) in OpenCV convention. Rows of R: [right, down, forward].

Average all numeric values across a list of dicts. Add num_scenes.

Calculate the Area Under the Curve (AUC) for the given error arrays.

Implementation of the forward pass through the DPT head. This method processes a specific chunk of frames from the sequence.

Get currently selected ViewerData.

List and filter image paths from a folder using args filters.

Return methods for a dataset/scene (pre-built at startup, gt always first).

Linearly interpolate between two cameras.

Dump a one-line summary of the aggregator's KV cache occupancy. Prints FlashInfer manager stats (if present) and SDPA dict stats otherwise, p

Rebuild playback controls.

Save batch results to JSON and return the number of failures.

Set a nested attribute via dotted path, e.g. 'scene.voxel_size'.

Update trajectory visibility based on checkbox.

Run `passes` full `clean → Phase-1 → N-stream` sequences. When ``kf_int > 1`` the inner stream loop alternates keyframe /

Apply resize transform to an image. Determines image dimensions automatically and computes the transform if not yet cached.

Apply resize transform with nearest-neighbor interpolation. Suitable for depth maps, masks, and integer-valued arrays. Args:

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

Build octree + centroid hierarchy from raw unprojected points. Args: all_xyz: (N, 3) float32 CPU tensor, world coordinates.

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

Remove directory contents only if not yet marked complete.

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

Return metadata dict to be merged into .complete.json. Returns: Dict with 'frame_keys', 'global_keys' lists (sorted), and optiona

Get configuration for a specific dataset. Returns: Dictionary with: - 'dataset_class': Dataset module name (e.g., 'se

Get selected dataset names from the 'datasets' list in base config. Returns an empty list if 'datasets' is absent or empty. Raises Ke

Compute (or return cached) transform for the given input dimensions. Args: orig_w: Original image width orig_h: Origi

Perform ICP registration to align source points to target points. This function implements Iterative Closest Point (ICP) algorithm using Open3D.

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

Invert 4x4 SE3 transformation matrix using R^T (assumes orthogonal R). Uses the efficient SE3 inverse: [R^T | -R^T @ t] which is valid when R

List image files in a folder.

Load a dataset class by name. Searches for the class in the datasets/ directory. Args: name: Dataset name (e.g., 'seven_

Load image from EXR format with adaptive channel support. Args: input_file: Input EXR file path return_channels: How to return th

Load NPZ produced by batch_demo.py. Accepts either: - A single ``.npz`` file path - A directory of per-frame ``frame_*.npz`` files (p

Select visible points via octree LOD traversal. Traverses the octree from root to leaves. At each level, cells are classified as:

Parse color string → [r, g, b] float 0-1. Accepts: '#RGB', '#RRGGBB', english name, single letter. Returns None for empty string.

Parse and filter point cloud data.

Convert predictions to the unbatched NumPy format used by vis code.

Return (TrailStyle, HeadStyle or None) for a named preset.

Render current view and return JPEG bytes.

Save image to EXR format with adaptive channel support. Args: image: Image array in one of the following formats: - HxW (sing

Save frame data for all frames in a scene (with parallel I/O). Frame files are named using zero-padded 6-digit indices (e.g., 000000.png),

Save global scene-level data. Args: global_data: Dictionary from load_global_data()

Save point cloud to binary PLY format. Args: points: Numpy array with shape [..., 3/4/6/7] - [..., 3]: xyz coordinates only

Plot AUC and accuracy metrics vs number of frames and save to output_dir. Args: frame_counts: Array of frame counts a

Segment a batch of RGB images. Returns list of bool arrays (True = sky).

Perform Umeyama registration to align source points to target points. This function implements the Umeyama algorithm for estimating a similarity

Batch GPU unproject. Returns (pts_xyz, pts_rgb, valid_counts) on CUDA.

Initializes the position generator with an empty cache.

Compute scale using median ratio.

Applies 1D rotary position embeddings along one dimension. Args: tokens: Input token features. positions: Position in

Apply positional embedding to tensor x.

Build camera segments from legacy flat CLI args. Handles the old --birdeye_start/--birdeye_duration/follow params. If no birdeye args, return

Collect sorted image files from a directory.

Compute mean optical flow magnitude induced by camera motion. Projects current frame pixels into the last keyframe camera using the current d

Compute mean optical flow magnitude induced by camera motion. Projects current frame pixels into the last keyframe camera using the current d

Compute scene center and scale from camera positions and point clouds. Returns: Tuple of (center as 3D array, scale as float dist

Deep-merge two dicts; override values take precedence. Args: base: Base dictionary override: Override dictionary (values take

Execute the eviction that was deferred during the last forward pass.

Execute the eviction that was deferred during the last forward pass.

Return the first existing pose file among known candidates.

Compute follow-camera eye/center/up for one frame. Returns (eye, center, up, forward_h) — forward_h is carried to next frame as fallback when

Return the cache directory for the current data selection, or None.

Return sorted list of frame base names (without extension) from custom_undistorted_cam/. Frames listed in DA3_FILTER_KEYS[scene] are excluded

Get rainbow gradient color based on progress (0.0 to 1.0). Color progression: Red -> Orange -> Yellow -> Green -> Cyan -> Blue -> Pur

Return sorted image paths, extracting video frames if needed.

Get 2D position embeddings for RoPE. Args: B: Batch size S: Sequence length H: Image height

Return scenes for a dataset (pre-built at startup).