Load NPZ data.

4×4 T_cam0_lidar: lidar frame → cam0 frame.

Return K (3×3), dist (4,), W, H from cam0.yaml.

Load all confidence maps in frame index order. Returns: List of HxW confidence maps (float32), resized if resize_context is set.

Convenience: load from --config YAML if given, then merge CLI args.

Load evaluation results from eval/{eval_type}.json. Returns None if not found.

KITTI Odometry does not provide a GT global point cloud.

Load global point cloud from points.ply. Returns: Nx3 or Nx6 array of global points, or None if not available

Load images from folder or video and preprocess into a tensor. Returns: (images, paths, resolved_image_folder): preprocessed tensor, file

Load and preprocess frames from video, with optional frame caching. If ``save_frames_dir`` is set: - First check if frames are already saved

Build GCTStream with random weights (no checkpoint). Eval mode on device.

Load GCTStream model from checkpoint.

Load cached sky masks or generate them with the ONNX model. Args: batch_size: Number of images to process per ONNX batch inference c

Load point cloud from binary PLY file (Open3D compatible). Args: input_file: Input PLY file path Returns: Numpy array with s

Load all point cloud maps in frame index order. Returns: List of HxWx3 or HxWx6 point cloud maps (float32), or None if not availa

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

Load RGB thumbnails from the npz cache.

Load NPZ data and build point clouds.

Load per-frame binary sky masks from cache.

Load full-resolution TLS PCD. Returns: pts : (N, 3) float32 — XYZ coordinates colors : (N, 3) float32 in [0, 1] — RGB, or None

Load trajectory alignment transform from traj_transform.txt. Returns: 4x4 Sim(3) transformation matrix, or None if file does not

Parse a TUM-format trajectory file. Returns: timestamps : (N,) float64 lines : list[str] — original text lines (no trailing

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

Generate a static bird's-eye CameraPath (single keyframe).

Blend depth colormap with the rectified image. Valid depth pixels show the TURBO colormap; invalid pixels show the original image.

Pivot camera: fixed eye, lookat follows scan trajectory direction.

Static camera: fixed eye and lookat for the entire segment.

Merge argparse namespace into config. Returns new config (no mutation). Supports both flat CLI args (--voxel_size) and dotted (--scene.voxel_

Reload scene data when Load is clicked.

Rotate around look_at point. dx = horizontal, dy = vertical.

Pan camera in screen plane.

Plot a precision-recall CDF curve and save to file. Ported from the official TAT toolbox plot_graph(). Red curve = precision (pred→GT distanc

Convert 2D position grid (HxWx2) to sinusoidal embeddings (HxWxC) Args: pos_grid: Tensor of shape (H, W, 2) containing 2D coordinate

Convert pose encoding to extrinsics (c2w) and move to CPU.

Converts GCT predictions to a 3D scene represented as a GLB file. Args: predictions: Dictionary containing model predictions with ke

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

Prepare a single scene. Args: gt_artifact: BSSArtifact for the GT directory scene: Scene name (e.g. 'chess/seq-01')

Side-by-side FPS / ms table across all variants.

Process rgb_dir: read images, predict sky masks, write PNGs to sibling sky_mask/.

Process one sequence. For every matched frame: 1. Decode & rectify the cam0 image → save images/{:06d}.png 2. Save the matched C

Process one scene: inference, optional NPZ / GLB, video render. Args: video_images: Pre-loaded video frames tensor (if from --video_path)

Run streaming inference. Return (per_frame_ms, scale_frames, phase1_ms). Reuses a single CUDA event pair across frames and syncs after every

pinhole camera projection

map points from ii->jj

Read and organize point cloud data.

Read camera intrinsics file in BSS v2 format. Format (one line per frame): frame_idx fx fy cx cy width height Args: intr_fil

Read trajectory file in BSS v2 format. BSS Trajectory Format v2 (one line per frame): frame_idx r00 r01 r02 tx r10 r11 r12 ty r20 r21 r22

Render one frame: LOD select -> overlays -> render -> return (H,W,3) uint8.

Render pre-culled points + extra geometry. No Scene needed. Used by D-lite parallel workers that receive visible points via SharedMem

Render one NPZ file to video / GLB.

Render high-res screenshot and save as PNG.

Calculate rotation angle error between ground truth and predicted rotations. Args: rot_gt: Ground truth rotation matrices ro

Run model inference and return visualization-ready numpy predictions.

D-lite parallel rendering. Producer (main): GPU cull → gather visible points → SharedMemory (exact size) Consumers (workers): SharedMemory →

Run method on a single scene in the current process. Returns None on success, Exception on failure.

Run method on a single scene. Returns None on success, Exception on failure.

Runs sky segmentation inference using ONNX model. Args: onnx_session: ONNX runtime session input_size: Target size for model

Run ONNX sky segmentation on a batch of BGR images. Returns list of uint8 score maps.

Dispatch a (method, dataset[, scene]) job to run_worker.py via conda run. Returns the subprocess exit code.

Run viser server for benchmark visualization. Args: workspace: Path to workspace directory. port: Port number for viser server.

Save binary mask image. Args: mask: HxW boolean mask array (True for valid) output_path: Output file path

Save threshold-colored PLY files and PR curves to vis_dir. Files are named with tau in the filename: vis_dir / f'gt_points_{tau}.

Save predictions as per-frame .npz files in a directory (parallel I/O). Creates a directory at ``output_path`` (without extension) and writes one

Save RGB image. Args: image: HxWx3 RGB image array (uint8, 0-255) output_path: Output file path