Sparse tensor with support for both torchsparse and spconv backends. Parameters: - feats (torch.Tensor): Features of the sparse tensor.

Stores the pose of an object. Also, stores some information about the scene that was used to normalize the pose.

This is the canonical representation of pose targets, used for computing metrics. instance_pose <-> invariant_pose_targets <-> all other

Embeds spatial positions into vector representations.

Renderer for the Voxel representation. Args: rendering_options (dict): Rendering options.

Embeds scalar timesteps into vector representations.

Handles remapping of 3D point coordinates and their inverse transformations.

A 3D subdivide block that can subdivide the sparse tensor. Args: channels: channels in the inputs and outputs. out_channels:

A GroupNorm layer that converts to float32 before the forward pass.

Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.

Sparse Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.

Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.

Renderer for the Voxel representation. Args: rendering_options (dict): Rendering options.

Preprocessor configuration for image, mask, and pointmap transforms. Transform application order: 1. Pointmap normalization (if normaliz

Sparse Transformer block (MSA + FFN).

Embeds scalar timesteps into vector representations.

Sparse Voxel Octree (SVO) implementation for PyTorch. Using Depth-First Search (DFS) order to store the octree. DFS order suits rendering

Fusing individual condition embedder. Require kwargs for the forward! Args: embedder_list: List of Tuples. Each Tuple consists of a c

Identity convention - no transformation applied. Direct passthrough mapping between instance pose and pose target values. This preserves

Sparse Transformer block (MSA + FFN) with adaptive layer norm conditioning.

Transformer block (MSA + FFN) with adaptive layer norm conditioning.

x_instance_scale and x_translation_scale are normalized to x_scene_scale

x_instance_scale and x_translation_scale are normalized to x_scene_scale x_instance_translation is unit

Projects (x,y,z) → D Splits into patches (patch_size x patch_size) Runs a tiny self-attention block inside each window Returns one to

Converts pose_targets <-> instance_pose <-> invariant_pose_targets

RandomResizedCrop for matching TF/TPU implementation: no for-loop is used. This may lead to results different with torchvision's version.

Midas eq. (6): https://arxiv.org/pdf/1907.01341v3 But for pointmaps (see MoGe): https://arxiv.org/pdf/2410.19115

Midas eq. (6): https://arxiv.org/pdf/1907.01341v3 But for pointmaps (see MoGe): https://arxiv.org/pdf/2410.19115

Downsample a sparse tensor by a factor of `factor`. Implemented as average pooling.