nn.Module wrapper around Tensor.permute for cleaner nn.Sequential usage.

ConvNeXt-style block adapted for 1D sequences (Temporal). Can use Attention, MLP-Mixer, or 1D Convolutions.

Standard ConvNeXt Block for 2D spatial processing.

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

Encodes motion information by jointly processing: - Correlation volume (matching cost) - Current flow estimate (2D / 3D / high-dim pa

Registry that ensures all datasets are auto-discovered and imported on first use.

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

Block containing Self-Attention and MLP.

Stores configuration for ConvNeXt blocks.

Correlation Block for RAFT-like architectures. Computes all-pairs correlation between fmap1 and fmap2 and builds a pyramid. Allows lookin

The Aggregator applies alternating-attention over input frames, as described in VGGT: Visual Geometry Grounded Transformer. Args:

Pads images such that dimensions are divisible by a certain stride (default 64). Useful for architectures that require specific input resolu

Block containing Cross-Attention and MLP.

Main application class for Depth Anything 3 Gradio app.

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

Prepares a batch of images for model inference. This processor converts a list of image file paths into a single, model-ready tensor. The proc

Linear head for dust3r Each token outputs: - 16x16 3D points (+ confidence)

Beautiful tabular metrics printer with color support. Features: - Colorized best/worst values - Grouped by dataset and evaluation mo

MLP as used in Vision Transformer, MLP-Mixer and related networks

2D image to patch embedding: (B,C,H,W) -> (B,N,D) Args: img_size: Image size. patch_size: Patch token size. in_chans

return positions of patches

Standard ResNet Residual Block.

Standard Residual Convolutional Block with GroupNorm or LayerNorm. Consists of: Norm -> Activation -> Conv -> Norm -> Activation -> Conv.

Lightweight residual convolution block for fusion

Dataclass for storing video tracks data.

Standard Multi-Head Attention.

Dataset for loading BlinkVision flow data. Loads RGB images, depths, masks, and flow particles from .npz files.

ConvNeXt Backbone implementation. Can initialize weights from torchvision's pre-trained models.

Main evaluation orchestrator for DepthAnything3 benchmarks. Usage: evaluator = Evaluator( work_dir="./eval_workspace",

Handles all event callbacks and user interactions for the Gradio app.

Top-down fusion block: (optional) residual merge + upsampling + 1x1 contraction

Handles file uploads and processing for the Gradio app.

3DGS parameters, all in world space

Decoder Head for the Track4World model. Handles upsampling of features extracted from the backbone and projects them to generate geomet

Response model for inference API.

Unified inference service class

Dataset loader for KITTI Scene Flow data stored in .npz format.

Dataset loader for Kubric 3D Flow data stored in .npz format.

Model backend service with persistent model loading.

Handles model inference and data processing for Depth Anything 3.

Output processor for converting model outputs to Prediction objects. Handles tensor-to-numpy conversion, batch dimension removal, and cr

Stable PCA→RGB for video features shaped (T, H, W, D) or a single frame (H, W, D). - Global mean/V3 reference for stable colors - Per-fra

Generates and caches 2D spatial positions for patches in a grid. This class efficiently manages the generation of spatial coordinates for patches

Generates and caches 2D spatial positions for patches in a grid. This class efficiently manages the generation of spatial coordinates for patches

Relative Update Block for 3D flow / scene motion. Key characteristics: - Relative refinement of 3D motion fields - Multi-sou

Relative Update Block for 2D flow refinement. Core idea: - Update flow in a *relative* manner rather than absolute prediction

1x1 convolution residual block

2D Rotary Position Embedding implementation. This module applies rotary position embeddings to input tokens based on their 2D spatial positio

2D Rotary Position Embedding implementation. This module applies rotary position embeddings to input tokens based on their 2D spatial positio

Task status model.

Track4World Model Implementation. Combines DINOv2 backbone for geometry estimation with a flow estimator (RAFT-like) for 4D (3D + Time) r

Dataset class for loading tracking evaluation data (ADT, DS, PO, etc.). Loads .npz files containing images, tracks, and camera parameters.

Handles UI component creation and layout for the Gradio app.

Dataclass for storing video tracks data.

Handles visualization updates and navigation for the Gradio app.

Basic ResNet-style encoder for feature extraction.

GRU-like update block used in RAFT-style iterative flow refinement. Combines: - Previous flow features - Context features

Standard Transformer block consisting of: - LayerNorm + Multi-Head Self-Attention - LayerNorm + Feed-Forward Network (MLP)

CameraHead predicts camera parameters from token representations using iterative refinement. It applies a series of transformer blocks (the

Clean and stabilized update block with LayerScale. Design goals: - Minimal, symmetric architecture - Stable deep refinement