Encode HD RGB video from a folder of original-resolution frames.

Export current filtered point clouds and cameras as a GLB file.

Extract frames from video using ffmpeg, cached in a sibling dir.

Gather all visible K and V tokens from the paged cache into dense tensors. Used by force_fp32 mode to bypass the FlashInfer FA2 k

Generate point clouds using loader's load_point_cloud_grid or global points. Args: loader: BSSLoader instance spatial

Generate 3D RoPE positions for streaming mode with correct global frame indices. Args: num_frames: Number of frames in c

Lazy-load COLMAP results (run pipeline if not cached).

Lazy-load and cache all COLMAP poses for a scene.

Lazily initialize FlashInferKVCacheManager on first use. Args: device: Device for cache tensors. dtype: Data type for

Initialize 3D RoPE for streaming inference.

Initialize frame_blocks and global_blocks from DINOv2 pretrained weights. Args: dino_ckpt: Checkpoint dictionary from DI

Initialize KV cache for streaming inference.

Load depth map. 7Scenes depth maps are 16-bit PNG in millimeters, need conversion to meters. Args: depth_file: Path to d

Read focal length from focal.txt. Args: scene: Scene name. Returns: Focal length in pixels (fx = fy).

Load and cache intrinsics from intrinsics.txt. File format: "fx fy cx cy width height" (single line). Returns [fx, fy, cx, cy], dtype

Load LingbotMap (GCTStream) model from checkpoint.

Load a per-frame NPZ directory with parallel I/O. Returns a single dict where each key has arrays stacked along a new leading dimension (S, .

Load pose file (4x4 C2W matrix). Args: pose_file: Path to pose text file Returns: 4x4 C2W transformation mat

Load RGB image. Args: color_file: Path to RGB image file Returns: HxWx3 RGB image (uint8, 0-255)

Load trajectory, returns dict mapping frame index (int) to C2W matrices. Args: loader: BSSLoader instance Returns:

W2C (4x4) in OpenCV convention. Camera +Z = forward.

Create viser handles for a single frame.

Create camera trajectory as gradient-colored cylinders. Args: positions: (N, 3) camera positions in world space radius: cylinder

Move viewer camera to match reconstructed camera at given frame.

Return the index of the nearest timestamp, or None if the array is empty.

Orthogonalize the rotation part of a 4x4 SE3 matrix via SVD. Required because evo validates SO(3) membership, but some datasets (e.g., 7Scene

Compute (scale, R, t) that maps *curr* into *prev*'s coordinate frame. Anchors on the latest physical time step inside the overlap region

Compute (scale, R, t) that maps *curr* into *prev*'s coordinate frame. Anchors on the latest physical time step inside the overlap region

Parse a TUM groundtruth.txt into timestamps and C2W matrices. Format per line: timestamp tx ty tz qx qy qz qw (comment lines start with '#')

Read a 3x3 intrinsics matrix and return [fx, fy, cx, cy].

Extract [fx, fy, cx, cy] from the P2 row of KITTI calib.txt. P2 is a 3x4 projection matrix for the left color camera (cam_2), in the rectifie

Parse LINGBOT_DEBUG_KV env var into a print-every-N interval. Accepts: "" (disabled) | "0" (disabled) | "1" (every frame) | "N" (eve

Parse KITTI poses file: one line per frame, 12 floats = 3x4 row-major C2W.

Parse raw NPZ arrays into the standardized format. Handles both single-file and per-frame-dir data (same key structure). Returns dict: image

Parse a TUM trajectory file into (timestamps, C2W 4x4 matrices). Each data line: timestamp tx ty tz qx qy qz qw (scalar-last quaternion) Th

Parse a TUM-style trajectory; key each pose by its integer timestamp. Each non-comment line: timestamp tx ty tz qx qy qz qw The timestamp is

Paste camera view parameters from global clipboard.

we will divide the token sequence into the following format [1 latent frame] [1 latent frame] ... [1 latent frame] We use fle

Convert list of HxWx3 uint8 numpy arrays to [S, 3, H, W] tensor in [0, 1].

Prepare special tokens (camera, register, optionally scale). Subclasses implement mode-specific token preparation. Args:

Causal attention for streaming inference using FlashInfer KV cache. Args: tokens: Input tokens [B*S_local, P, C]

Process frame attention blocks. Frame attention operates independently per frame (no cross-frame communication). Tokens stay

Process global (cross-frame) attention blocks. Subclasses implement mode-specific attention logic. Args: tokens

Convert model predictions to benchmark output format. Args: predictions: Raw model outputs with 'pose_enc', 'depth', 'depth_conf'

Process prediction dictionary to extract visualization data. Args: pred_dict: Model prediction dictionary. use_point_

Read a COLMAP SfM .log file and return list of 4x4 C2W matrices. Log format: 5 lines per entry Line 1: metadata (idx idx 0) Lines 2-5

Find mask/ directory: check self.artifact.root first, then sibling gt/.

Resolve a raw config value into a concrete keyframe interval. ``None``, ``0``, or the string ``"auto"`` triggers auto-selection: ``1`` when `

Performs feature rotation by splitting and recombining feature dimensions. Args: x: Input tensor to rotate. Returns:

Compute rotation angle error between ground truth and predicted rotations.

Run COLMAP SfM pipeline and parse results. Returns: Dict with 'poses' and 'intrinsics' keyed by image filename, or No

Run LingbotMap inference and return raw predictions dict.

Handle main inference mode: load model, optionally compile, process scenes.

Handle --load_predictions mode: render or visualize saved NPZ files.

D-lite parallel rendering via SharedMemory.

Render all frames in the main process.

Handle --visualize_sky_mask_only mode for discovered scenes.

Run batched sky segmentation using ONNX-based SkySegmenter. Args: rgb_list: List of uint8 RGB images. Returns:

Minimal smoke test. Run with: python -c "from lingbot_map.layers.flashinfer_cache import _sanity_check; _sanity_check()"

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

Save original images as video.

Save global point cloud.

Setup GUI controls.

Setup camera token, register tokens, and optionally scale token. Subclasses implement mode-specific token initialization. M

Sort finest-level centroids by frame → CPU numpy for Scene.

Returns torch.sqrt(torch.max(0, x)) but with a zero subgradient where x is 0.

Returns torch.sqrt(torch.max(0, x)) but with a zero subgradient where x is 0.

Compute a static camera. Defaults from start frame pose if not given.

Concatenate per-window predictions while de-duplicating overlaps. For each temporal key the method builds a slice table first — every

Concatenate per-window predictions while de-duplicating overlaps. For each temporal key the method builds a slice table first — every

Capture a screenshot from the current view and save to file. Args: client: The viser client that triggered the action. If None,

Compute translation angle error between ground truth and predicted translations.

Update camera clipping planes based on scene bounds.

Update camera frustums visibility based on checkbox.

Update scene bounding box.

Add/update camera trajectory as gradient-colored line segments.