Initialize LearnedAlignment. Args: steps: Number of steps in the UNet. stride: Effective downsampling of the alignmen

Initialize MonodepthFeatureExtractor.

Initialize Sliding Pyramid Network. The framework 1. creates an image pyramid, 2. generates overlapping patches with

Initialize ViT from TIMM implementation.

Initialize UNet Encoder. Args: dim_in: The number of input channels. width: Width multiplicator of intermediate layer

Initialize UNet Decoder. Args: dim_out: The number of output channels. width: Width of last input feature map from en

Initialize multiresolution convolutional decoder. Args: dims_encoder: Expected dims at each level from the encoder. d

Create multiple disparity layers.

Create multiple surface layers.

Add a single frame to output.

Convenience method to convert the class to a dict.

Compute gradient of clamp with pushback.

Finish writing.

Colorize alpha map.

Configure logger globally. Args: log_level: The desired verbosity level. log_path: The path to write logs to. prefix: The

Create activation function and corresponding inverse function. Args: activation_type: The activation type to create. Returns:

Create camera model to simulate general pinhole camera.

Create eye trajectory for trajectory type.

Decode color space index to color space.

Encode color space to integer.

Predict images from gaussians. Args: gaussians: The Gaussians to render. extrinsics: The extrinsics of the camera to

Apply clamp.

Apply residual block.

Process and fuse input features.

Optionally align monodepth to ground truth with a local scale map. Args: monodepth: The monodepth model with intermediate feature

Predict 3D Gaussians. Args: image: The image to process. disparity_factor: Factor to convert depth to disparities.

Apply mean and std normalization over input image.

Apply affine range normalization over input image.

Construct Gaussian base values and prepare feature input. Args: image: The image to process. depth: The corresponding

Decode by projection and fusion of multi-resolution encodings.

Process image and return disparity and feature maps.

Apply SkipConvBackbone to image.

Run monodepth and fuse features with input image to predict Gaussians. Args: input_features: The input features to use.

Predict deltas for 3D Gaussians. Args: image_features: Image features from decoder. Returns: The predicted d

Compute alignment map.

Apply projection module.

Extract multi-resolution features.

Encode input image into multi-resolution encodings.

Encode input at multiple resolutions.

Override forwarding with intermediate features. Adapted from timm ViT. Returns: Output features and list of features fro

Apply UNet Encoder to image. Args: input: The input image. Returns: The output multi-level feature map from

Apply UNet to image. Args: features: The input multi-level feature map from encoder. Returns: The output fea

Decode the multi-resolution encodings.

Decode (multi-resolution) encodings.

Freeze all normalization layers.

Return supported image extensions.

Return supported video extensions.

Apply hard sigmoid with pushback. For compatibility reasons, we follow the default PyTorch implementation with a default slope of 1/6:

Return the internal image size of the network.

Internal resolution of the encoder.

Internal resolution of the encoder.

Return the internal image size of the network.

Compute inverse sigmoid.

Compute inverse softplus.

Check if a number is a power of two.

LinearRGB to sRGB conversion function. Reference: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf Section 7.7.

Load an RGB image.

Run inference for SHARP model.

Compute the output width for UNet decoder.

Output resolution of Gaussians.

Predict Gaussians from input images.

Project tensor u to unit basis a.

Get conjugate of a quaternion.

Compute dot product between two quaternions.

Convert batch of rotation matrices to quaternions. Args: matrices: The matrices to convert to quaternions. Returns: The quat

Compute relu with pushback.

Predict Gaussians from input images.

Convert batch of quaternions into rotations matrices. Args: quaternions: The quaternions convert to matrices. Returns: The r

Save image to given path.

Enable grad checkpointing.

Enable grad checkpointing.

Modify the default extrinsics.

Clamp tensor to min/max in differentiable way. Args: tensor: The tensor to clamp. min: Pair of threshold to start clamping and va

Effective downsampling stride.

Internal stride of GaussianDensePredictionTransformer.

Write rendered color/depth/alpha to files.