Convert W2C (rotation + translation) to 4x4 C2W matrix.

Apply a similarity transform (s, R, t) to one window's predictions.

Apply a similarity transform (s, R, t) to one window's predictions.

Allocate one free patch page and write patches_per_frame patch tokens. Direct tensor assignment to kv_caches[block_idx][page_id,

Append num_special_tokens (6) special tokens to the special stream. Direct tensor slice assignment to kv_caches[block_idx][tail_p

Process network output to extract 3D points and confidence values. Args: out: Network output tensor (B, C, H, W) activation:

action on point cloud

Add point cloud for a frame.

Read Layer 1 json files from disk, generate Layer 2 and Layer 3. Layer 2: per-scene cross-method comparison (read-modify-write) Layer 3: data

Average per-scene AUC/Racc/Tacc values.

Pool rError/tError from all scenes, compute AUC once.

Setup and run animation controls.

Apply range / stride / first_k / last_k filters to image paths.

Apply sampling strategy to frame list. Args: frames: Original frame ID list (integers) sampling_config: Sampling conf

Visibility filter (CMRNet / iralab visibility_kernel2). Removes depth pixels that are occluded by a foreground surface. Dispatches to th

Phase 2 streaming only: norm1 + prepare_qkv fused as one compilable unit. Extracted as a named method so torch.compile can capture norm1 + qk

Produce the final immutable Scene.

Fixed hardware calibration (sensor.yaml): t=[0,0,0.124 m], q_xyzw=[0,0,1,0] → 180° Z rotation + 124 mm Z lift. Lidar frame → base fr

Calculate how many frames to unproject in one GPU batch. Args: H, W: frame dimensions downsample: spatial downsample

Build indices for all possible pairs of frames. Args: N: Number of frames B: Batch size Returns: i1, i2: Indice

Compute Chamfer distance: average of accuracy and completeness. The Chamfer distance is a symmetric metric that combines both directions of n

Clean up resources when client disconnects.

Delete entire eval/ directory.

Return sorted list of image files in rgb_dir.

(H,W) float → (H,W,3) uint8 BGR colorized depth.

Turn a grayscale image into a color image. Args: x: Input grayscale image [H, W] cmap_name: Colormap name mask: Opti

Normalize the translation vectors and compute the angle between them. Args: t_gt: Ground truth translation vectors t: Predic

Apply torch.compile(mode="reduce-overhead") to compute-heavy, fixed-shape modules and drop point_head. Matches the optimizations from the ori

Compile hot, fixed-shape modules with mode="reduce-overhead". Mirrors the targets in gct_profile.py:compile_model. Unlike the profile script,

Apply torch.compile(mode='reduce-overhead') to model sub-modules. Follows the three-phase strategy from the reference profiling script: 1.

Compute how AUC changes as more frames are added. Args: gt_loader: BSSLoader for ground truth directory pred_loader

Computes faces for a camera mesh with multiple shells (for thicker frustums). Connects each consecutive pair of vertex shells to form the frustum

Diagonal of the first-N-frames trajectory bounding box.

Frame-boundary pointer array (same semantics as VoxelGridCUDA).

Diagonal of the camera trajectory bounding box.

Create a normalized UV grid of shape (width, height, 2). The grid spans horizontally and vertically according to an aspect ratio, ensuri

Convert a depth map to camera coordinates. Args: depth_map (np.ndarray): Depth map of shape (H, W). intrinsic (np.ndarray):

Convert a depth map to world coordinates. Args: depth_map (np.ndarray): Depth map of shape (H, W). intrinsic (np.ndarray): C

Downloads a file from a URL, handling redirects.

Download sky segmentation model from HuggingFace. Args: output_path: Path to save the model Returns: Path to the downlo

Side-by-side video: rendered point cloud (left) + RGB (right). Both streams are resized to the same height before concatenation.

Colorize depth frames (S,H,W) and encode to MP4.

Encode PNGs to MP4. Tries FFmpeg first, falls back to OpenCV.

Get expected dataset class name. Args: name: Dataset module name. Returns: Expected class name (PascalCase +

Get expected method class name. Args: name: Method module name. Returns: Expected class name (PascalCase + '

Load the optional CUDA visibility extension on first use.

Run all enabled evaluators, save results to individual eval/*.json files.

Optional method for custom point cloud evaluation. Args: gt_loader: BSSLoader for the ground truth directory (no resize_context

Strip leading date (YYYY-MM-DD-) and trailing sequence number (-N or -NN). '2024-03-12-keble-college-02' → 'keble-college'

Compute F1-score at multiple distance thresholds. F1-score is the harmonic mean of precision and recall, providing a balanced measure of reco

FFN residual branch: norm2 -> mlp -> ls2, WITH residual add fused in. Includes the residual add (x + ...) so torch.compile captures the entir

Nearest-neighbour search. Returns (index, diff) if diff <= max_gap, else (None, diff).

Find scenes under an input folder.

Frustum cull the TLS cloud and build a depth map for one frame. GPU path (pts_gpu is not None and _HAS_VIS_EXT): Calls the compiled

计算1D旋转位置编码（RoPE）的频率张量。 RoPE的核心思想：使用旋转矩阵来编码位置信息，使得相对位置关系保持不变。 公式：对于位置m和维度i，频率为 θ_i = θ^(-2i/d)，其中θ是基础频率（默认10000） Args:

Return workspace/{dataset}/eval/

Get the host IP address for the viser server.

Return Path to ONNX model. Raises FileNotFoundError if not present.

Return (xyz, rgb) for the requested view, filtered by confidence.

Get processing dimensions (what methods see after resize). Returns: (width, height) tuple

Return workspace/{dataset}/{scene_safe}/eval/

Return the TLS merged-cloud PCD path for *scene_kw*, or None if absent.

Create a vertical colorbar image. Args: h: Height in pixels vmin: Minimum value vmax: Maximum value cmap_nam

Return True if a sky mask cache file already exists.

Applies intrinsic parameters and optional distortion to the given 3D points. Args: cam_intrinsics (torch.Tensor): Intrinsic camera p

Interpolate keyframes with cubic spline. Returns list of (camera, frame_idx).

pinhole camera inverse projection

Check if cache exists and parameters match. Args: num_frames: Number of frames in the scene image_width: Width o

Check if the RGB thumbnail cache exists and matches current parameters.

List all methods that have output directories for a scene.