Process model results into structured data. Args: target_dir: Directory containing results prediction: Model

Process video file and extract frames. Args: input_video: Path to input video file target_dir_images: Target

X: (N,D) where N = H*W Returns PCs_raw: (N,3) using stable basis (fixed or Procrustes-aligned)

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

Run model forward pass.

Sample frames if scene has more than max_frames. Uses fixed random seed (42) for reproducibility. Args: scene_d

Save predictions data to predictions.npz for caching. Args: target_dir: Directory to save the cache predicti

Schedule task cleanup in background.

Set up all event handlers for the application. Args: demo: Gradio Blocks interface All other arguments: Grad

Set up example scene handlers.

Set up navigation handlers for measure tab.

Set up visualization update handlers.

Check if image should be filtered out based on known problematic views.

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.

Center-crop all tensors to the smallest H, W; adjust intrinsics' cx, cy accordingly.

results: List[Tuple[torch.Tensor, Tuple[H, W], Optional[np.ndarray], Optional[np.ndarray]]] -> processed_images, out_sizes, out_intri

Convert absolute pose sequence back to sequential relative transforms T_0, T_1, T_2, ... -> S_01, S_12, S_23, ...

Accumulate adjacent SIM(3) transforms into transforms from the initial frame to each subsequent frame. Args: transforms: list, each

Process network output to extract GS params and density values. Density could be view-dependent as SH coefficient Args: out: Ne

Run complete evaluation pipeline: inference + evaluation. Args: api: DepthAnything3 API instance Returns:

Apply metric scaling to depth based on camera intrinsics. Args: depth: Input depth tensor intrinsics: Camera intrinsics tens

PyTorch SIM3 point_maps: (b, h, w, 3) numpy array s: scalar or (b,) array R: (3, 3) or (b, 3, 3) numpy array t: (3,) or (b, 3) nu

Build list of groups from gallery directory.

Build manifest for a specific group.

Build loop closure constraints

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

Convert camera space points to world space points. Args: camera_space_points (torch.Tensor): Camera space points with shape (b, v, h

Args: camray: (B, S, num_patches_y, num_patches_x, 6) confidence: (B, S, num_patches_y, num_patches_x) Returns: R: (B

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

Compute mask for depth alignment based on confidence and depth thresholds. Args: depth_conf: Depth confidence tensor non_sky

method: 'auto', 'ransac', 'weighted'

Args: rays_origin (torch.Tensor): (B, N, 3) rays_target (torch.Tensor): (B, N, 3) z_threshold (float): Threshold for z va

Create the 3D viewer component. Returns: 3D model viewer component

Create the acknowledgements section. Returns: Acknowledgements HTML component

Create FastAPI application with model backend.

Create and configure the Gradio application. Returns: Configured Gradio Blocks interface

Create the description section. Returns: Description HTML component

Create the display control section (options for visualization). Returns: Tuple of display control components including b

Create the example scene grid. Args: scenes: List of scene information dictionaries Returns: List o

Create the example scenes section. Returns: List of scene information dictionaries

Create the header section with logo and title. Returns: Header HTML component

Create the inference control section (before inference). Returns: Tuple of (process_res_method_dropdown, infer_gs, ref_v

Create the measurement section. Returns: A tuple of (prev_measure_btn, measure_view_selector, next_measure_btn, measure_

Create the 3DGS rendered video display component and info message. Returns: Tuple of (video component, info message comp

Generate a ring of Sim3 transforms with rotation, adding slight rotational noise

Create the upload section with video, images, and gallery components. Returns: A tuple of Gradio components: (input_vide

Detect input type from path. Returns: - "image": Single image file - "images": Directory containing images - "video"

Heuristic estimate for memory usage (GB) based on image count and resolution. This mirrors the simple policy used by the backend service so other

Evaluate camera pose estimation accuracy. Args: scene: Scene identifier result_path: Path to .npz file conta

Export feature visualization with PCA. Args: prediction: Model prediction containing feature maps export_dir: Directory to export

Convert camera extrinsics and intrinsics to a compact pose encoding.

Fit global mean/V3 and initialize percentiles from a reference set. frames: ndarray (T,H,W,D) or list of (H,W,D)

Forward pass through the model. Args: image: Input batch with shape ``(B, N, 3, H, W)`` on the model device.

Fuse per-view depth maps into a single point cloud. Args: scene: Scene identifier result_path: Path to .npz

Generate a coordinate frame where the Y direction is normal to the plane defined by unit vectors a and b. The other axes are arbitrary.

Generate the acknowledgements section HTML. Returns: str: HTML string for the acknowledgements

Scans all YAML files in the configs directory and returns a sorted dictionary where: - Keys are model names (YAML filenames without the .yaml

Collect per-view image paths, intrinsics/extrinsics, and GT masks. Args: scene: Scene identifier (e.g., "scan1")

Generate the main description and getting started HTML. Returns: str: HTML string for the description

Get the configured Gradio theme with adaptive tech colors. Returns: gr.themes.Base: Configured Gradio theme

Generate the main header HTML with logo and title. Args: logo_base64 (str, optional): Base64 encoded logo image Returns: