Crops the padded tensor back to original dimensions.

1x1 convolution without padding

pack and write to a binary file. :param fid: :param data: data to send, if multiple elements are sent at the same time, they should be enc

Pads a list of input tensors.

Align estimated trajectory to reference using Umeyama Sim(3). Default no RANSAC; if ransac=True, use RANSAC (max iterations default 10).

Read and unpack the next bytes from a binary file. :param fid: :param num_bytes: Sum of combination of {2, 4, 8}, e.g. 2, 6, 16, 30, etc.

Perform a robust GPU cleanup sequence. This includes synchronizing, emptying caches, collecting IPC handles and running the Python garbage co

Accept (..., 3,4) or (..., 4,4) extrinsics, return (...,4,4) homogeneous matrix. Supports torch.Tensor or np.ndarray.

Create an object from config. The config is expected to contains the following: __object__: path: path.to.module name: MyClas

Evaluate for all configured modes and write JSON files. Evaluation order by mode (all datasets per mode): 1. pose -

Precompute RoPE coefficients.

Handle export directory

Process single image

Unified inference interface

Generate normalized UV coordinates on the camera view plane. Coordinates range: - Top-left: (-width / diagonal, -height / diagonal)

Uniformly sample points from a triangle mesh. Args: mesh: Input triangle mesh num_points: Number of points to sample Re

Remove ANSI escape sequences from string for length calculation.

Wait until file size stabilizes for 2 consecutive checks.

Generic function to draw text on a tensor image.

Validate path

Lift 3x3 matrices to homogeneous 4x4 matrices.

Converts [-0.5, 0.5] float to [0, 255] uint8.

Unified iterative flow update (RAFT-style). Handles both 2D and 3D flow updates using correlation volumes and recurrent units.

Return a snapshot of current GPU memory usage or None if CUDA not available. Keys in returned dict: total_gb, allocated_gb, reserved_gb, free_gb,

Run inference on input images. Args: image: List of input images (numpy arrays, PIL Images, or file paths) e

Initializes the Track4World model and loads pretrained weights.

Parse export_feat parameter

Simple UV position encoding directly added to feature map.

Apply activation to single channel output, maintaining semantic consistency with value branch in multi-channel case. Supports: exp /

Apply activation to single channel output, maintaining semantic consistency with value branch in multi-channel case. Supports: exp /

Closed-form Sim(3) estimation via SVD, aligning A to B. Returns: s, R, t

Get export directory path. Structure: .../model_results/{data}/{scene}/{posed|unposed}

Invert a 4x4 SE(3) matrix.

Solves the alignment problem: find 'a' that minimizes weighted L1 distance. If trunc is provided, it solves a Robust L1 problem using an exha

Align `points_src` to `points_tgt` with an isotropic scale and a 3D shift. The algorithm uses an 'anchor' strategy: 1. Centering the

Calculate the Area Under the Curve (AUC) for the given error arrays. Args: r_error: Rotation error values in degrees t_error

Create a scalable TSDF volume for depth fusion. Args: voxel_length: Size of each voxel sdf_trunc: Truncation distance for SD

Safe interpolation implementation to avoid INT_MAX overflow in torch.nn.functional.interpolate.

src_pts: (N,2) source points (torch.Tensor, float32/float64) dst_pts: (N,2) target points (torch.Tensor, float32/float64) confident_weigh

Sliding window inference for longer sequences. Optionally uses cached features (eval_dict) to speed up evaluation.

Fuse multiple depth maps into TSDF volume and extract mesh. Args: volume: TSDF volume to integrate into depths: Depth maps [

Convert batched points (xyz) to (xyz1).

Load a configuration. Will resolve inheritance. Supports both file paths and module paths (e.g., depth_anything_3.configs.giant).

Resize 2D map by nearest interpolation. Return the nearest neighbor index and mask of the resized map. ### Parameters - `inputs`: a

Compute nearest neighbor distances from verts2 to verts1 using KDTree. Args: verts1: Reference point cloud [N, 3] verts2: Qu

Converts a time-series feature [B, T, C, H, W] to [B*(T-1), 2, C, H, W] by stacking adjacent frames (t, t+1).

Print evaluation metrics in a beautiful tabular format. Args: metrics: Metrics dictionary. If None, loads from saved JSO

Estimate global focal (scalar) and shift (3D) per batch from video point maps. Args: points_all: Tensor (B, S, H, W, 3), point maps

Saves a mesh or point cloud to a .ply file using trimesh. Args: save_path: Destination path. vertices: (N, 3) array of verte

Scatter the minimum value along the given dimension of `src` into a new tensor. Unlike standard scatter, this also recovers the original indi

Simple UV positional embedding added to feature maps.

Accept (4,4) or (3,4) extrinsic parameters, return (4,4) homogeneous matrix.

Generic evaluation loop shared across datasets. Args: model: The model to evaluate. loader: The dataloader. batc

Join URL parts safely.

Back-project depth map into 3D world coordinates. Args: depth: Depth tensor [B, 1, H, W] proj: Projection ma

Get list of scenes to evaluate, optionally filtered.

Scan and import all .py submodules under depth_anything_3.bench.datasets (skip files/packages starting with underscore), to trigger @REGISTRY

Pads a cumsum tensor with a 0 at the start and repeats the last value.

Pads a tensor with -inf at the start and +inf at the end.

Prepare transforms for PRoPE-style positional encoding.

Process the next task in the queue.