Move Gaussians to device.

Helper function that applies checkpointing. If enabled applies grad checkpointing, otherwise just executes the function normally.

Create a simple 2D residual block.

Merge the patched input into a image with sliding window.

Create normalization layer.

Split the input into small patches with sliding window.

Compute disparity quantiles for scene and extrinsics id.

Colorize a scalar map of. Args: scalar_map: Map of with format BHW. val_min: Minimu value to display. val_max: Maximum va

Create batch of identity matrices.

Discard class token and reshape 1D feature map to a 2D grid.

Initialize MeanStdNormalizer.

Create a residual block.

Variant of clamp function which avoid the vanishing gradient problem. This function is equivalent to adding a regularizer of the form pu

Factory function for creating a ViT model.

Combine predicted delta values with base gaussian values and apply activation function. Args: delta: The delta values predicted b

Render a single gaussian checkpoint file.

Robust torch.where function to avoid NaN in backward pass. See https://github.com/pytorch/pytorch/issues/68425 Args: condition: When

Initialize ResidualBlock.

Initialize DepthAlignmentWrapper. Args: scale_map_estimator: Module to align monodepth to ground truth depth.

Initialize MonodepthWithEncodingAdaptor. Args: monodepth_predictor: The monodepth model. return_encoder_features: Whe

Initialize SkipConvBackbone.

Initialize projection module.

Convert conics to covariance matrices.

Create base scale for the gaussians.

Create base x and y coordinates for the gaussians in NDC space.

Create image encoder and return based on parameters.

Creates a 3-level image pyramid.

Return constants for scale activation function.

Mean activation function. Args: base: Tensor of shape [B, 3, H, W], where first two feature dimensions (x,y) are

Rescale a depth image tensor. Args: depth: The depth tensor to transform. depth_min: The min depth to scale depth to. dep

Add a single frame to output.

Apply an affine transformation to 3D Gaussians. Args: gaussians: The Gaussians to transform. transform: An affine transform with

Finish writing.

Colorize depth map.

Compose 3D covariance matrices into quaternions and singular values. Args: quaternions: The quaternions describing the principal basis.

Compose rendered RGB with background color.

Compute camera for eye position.

Compute the maximum offset for camera along X/Y/Z axis.

Converts a focal length given in mm to pixels.

Convert RGB to degree-0 spherical harmonics. Reference: https://en.wikipedia.org/wiki/Table_of_spherical_harmonics

Convert degree-0 spherical harmonics to RGB. Reference: https://en.wikipedia.org/wiki/Table_of_spherical_harmonics

Create depth alignment.

Create camera matrix from vectors.

Create a rotating trajectory.

Create a rotating trajectory.

Create a left right shake followed by an up down shake trajectory.

Create a left to right swipe trajectory.

Create gaussian_decoder model specified by gaussian_decoder_name.

Create inpainter.

Create an adaptor that returns both disparity and features.

Create DepthDensePredictionTransformer model. Args: patch_encoder_preset: The preset patch encoder architecture in SPN. dims_deco

Creates DepthDensePredictionTransformer model. Args: params: Parameters of monodepth network. Returns: The configured monode

Creates DepthDensePredictionTransformer model. Args: patch_encoder_preset: The preset patch encoder architecture in SPN. image_en

Create gaussian predictor model specified by name.

Decompose 3D covariance matrices into quaternions and singular values. Args: covariance_matrices: The covariance matrices to decompose.

Return exif information as a dictionary.

Generate cross product matrix for vector exterior product.

Return dimensions of output feature maps.

Get resolution for metadata dictionary.

Compute unprojection matrix to transform Gaussians to Euclidean space. Args: extrinsics: The 4x4 extrinsics matrix of the camera view.

Internal resolution.

Return the internal image size of the network.

Return the full image size of the SPN network.

Loads a ply from a file.

Predict Gaussians from an image.

Prepare the feature input to the Guassian predictor.

Replicate the last convolution layer (head[4] in DPT) for multi layer depth.

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

Save a predicted Gaussian3D to a ply file.

Enable grad checkpointing.

Set requires grad for separate components.

Unproject Gaussians from NDC space to world coordinates.

Upsample the delta value. Args: delta: The delta values predicted by gaussian predictor. scale_factor: The scale fact

Create upsampling layer.

Write image to binary stream.

Initialize VideoWriter.

Initialize gsplat renderer. Args: color_space: The color space to use for rendering. background_color: The background

Initialize GeneralPinholeCameraModel. Args: scene: The scene to display. screen_extrinsics: Extrinsics of the default

Initialize feature fusion block. Args: dim_in: Dimensions of input. dim_out: Dimensions of output. batch_

Initialize RGBGaussianPredictor. Args: init_model: A model mapping image and depth to base values. monodepth_model: T

Initialize AffineRangeNormalizer.

Initialize MobileNetNormalizer.

Initialize MultilayerInitializer. Args: stride: The downsample rate of output feature map. base_depth: The depth of t

Initialize Dense Prediction Transformer. Args: encoder: The SlidingPyramidTransformer backbone. decoder: The Multires

Initialize Dense Prediction Transformer for Gaussian. Args: decoder: Decoder to decode features. monodepth_decoder: O

Initialize DirectGaussianPredictor. Args: feature_dim: Number of input features. num_layers: The number of layers of

Initialize GaussianComposer. Args: delta_factor: Multiply delta offsets by this factor. min_scale: The minimal scale