Compute the inverse of each 4x4 (or 3x4) SE3 matrix in a batch. Uses closed-form solution instead of torch.inverse() for numerical stability

Compute non-sky mask from sky prediction. Args: sky_prediction: Sky prediction tensor threshold: Threshold for sky classific

Convert general model state dict to match current model architecture. Args: state_dict: Original state dictionary Returns:

Evaluate 3D reconstruction quality using standard metrics. This function computes: - Accuracy: Mean distance from predicted points to GT

Generate a coordinate frame given perpendicular, unit-length Y and Z vectors.

Extract feature maps for all frames. Processes frames in chunks to manage memory usage.

Get information about scenes in the examples directory. Args: examples_dir: Path to examples directory Returns: List of

Reads frames from a directory of images or a video file. Args: input_path: Path to image directory or video file. resize_to:

Solve jointly for optimal focal length and global depth shift. Solve min ||focal * (xy / (z + shift)) - uv||^2 using Levenberg-Marquardt.

Solve for optimal depth shift given a fixed focal length. Solve min ||focal * (xy / (z + shift)) - uv||^2 Args: uv: pixel coord

Splits a batch-processing function into smaller chunks to prevent OOM (Out of Memory).

Writes a video/gif to TensorBoard.

Summarizes a single image.

Update measure view for a specific view index. Args: processed_data: Processed data dictionary view_index: I

Compute weighted mean of a tensor.

Single block of MLP-Mixer.

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

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

Clean up old completed/failed tasks to prevent memory buildup.

Helper to convert tensor types.

For each frame, transform (u,v,1) through K^{-1} to get rays, multiply by depth to camera frame, then use (w2c)^{-1} to transform to world fr

Write JSON with UTF-8 and pretty indentation.

Compute reconstruction metrics for each dataset and scene. Args: mode: "recon_unposed" or "recon_posed"

4-layer top-down fusion, returns finest scale features (after fusion, before neck1).

Invert 3x3 intrinsics matrices. Assumes no skew.

Check if name is safe for use in paths.

GPU-accelerated nearest neighbor distance computation. Args: query: Query points [N, 3] target: Target point

Load GT depth and create valid mask. For ScanNet++, GT depth is stored as 16-bit PNG in millimeters. Returns: B

Load GT depth and create valid mask. For 7Scenes, GT depth is stored as 16-bit PNG in millimeters. Value 65535 indicates inv

Load GT depth and aliasing mask to create valid mask. For HiRoom: - GT depth is stored as 16-bit PNG, scaled to 100m range

Load GT mask for masking invalid regions. GT mask marks occluded or invalid regions that should be excluded from depth fusio

Mask invalid depth values by setting them to 0.

Mask invalid depth values by setting them to 0. Args: depth: Depth map to mask gt_zero_mask: Optional GT mas

Mask invalid depth values by setting them to 0.

Mask invalid depth values by setting them to 0. Logic: 1. Apply GT mask (if provided) - marks occluded/invalid regions

Compute elementwise mean across a list of homogeneous metric dicts.

Normalize extrinsics

Print report footer.

Print report header.

Processor for Kubric and BlinkVision dataset structures.

Read point cloud from PLY file.

Run model forward pass.

Save GT extrinsics/intrinsics/image_files for evaluation. This is needed when frames are sampled, so eval_pose and fuse3d can use

Common implementation for saving images synchronously or asynchronously

Convert numpy scalars to plain Python floats for JSON safety.

Align all camera poses to the first camera's coordinate frame. Args: camera_poses: Camera poses as SE3 transformations [N, 4, 4]

Asserts that two tensors have the same shape. Useful for debugging element-wise operations.

Wrapper around grid_sample with coordinate normalization. Args: input: Input feature map (N, C, H, W) or (N, C, D, H, W).

Pre-cache all example scenes at startup. Args: show_cam: Whether to show camera in visualization filter_blac

Arrange images in a line. The interface resembles a CSS div with flexbox.

Return whether at least ``required_gb`` seems available on the current GPU. The returned tuple is (is_available, message) with a human-friendly m

Apply color to a metric value based on whether it's best/worst. Args: value: String representation of the value is_best: Whe

Compute pose estimation metrics between predicted and ground truth trajectories. Args: pred_se3: Predicted SE(3) transformations [N,

Lightweight 3-layer CNN encoder (1x1 -> 3x3 -> 3x3), used to compress correlation features.

Convert metric model state dict to match current model architecture. Args: state_dict: Original metric state dictionary Returns

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

Returns val if it exists, otherwise returns default d.

DINOv2 ViT-L/14 model with registers (optionally) pretrained on the LVD-142M dataset.

Applies stochastic depth by computing residuals only on a subset of batch samples and scaling them accordingly. This reduces computation

Estimate a global Sim(3) transformation that aligns A to B in a least-squares sense. That is, find s, R, t such that: B ~ s * R * A +

Evaluate 3D reconstruction quality against ground truth. Args: scene: Scene identifier fuse_path: Path to fu