Loads data from a BSS artifact directory.

Single camera state in world coordinates.

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

Manages BSS hierarchy: workspace / dataset / scene / method. Responsible only for: path construction at that level, scene/method listing, and

Keyframe-based camera path with interpolation. Presets (follow, birdeye) are just functions that generate CameraPath. JSON format: {

Streaming GCT model with KV cache for efficient online inference. Features: - AggregatorStream with KV cache support (FlashInfer backend

Interactive orbit camera with WASD FPS-style movement.

Trajectory trail line style.

Manages configuration loading and parameter routing. Supports the split-directory layout: - configs/base.yaml : workspace + evaluat

Camera head visualization style.

Residual convolution module.

Streaming causal aggregator with FlashInfer paged KV cache. Features: - Temporal causal attention (each frame only attends to past frame

Manages saving of dataset/method frame and global data to BSS format. Tracks which frame and global keys were saved during the current session,

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

DPT Head for dense prediction tasks. This implementation follows the architecture described in "Vision Transformers for Dense Prediction"

Interactive 3D point cloud viewer with camera visualization. Features: - Point cloud visualization with confidence-based filtering -

Encapsulates a single resize/crop transformation. Attributes: original_size: (W, H) of the original image intermediate_size:

Path template for a single BSS directory (gt or method output). All well-known paths are fixed properties defined at __init__ time, making th

Configurable camera visualization: trail line + head display.

Two-stream paged KV cache: patch pages (recyclable) + special pages (append-only). Args: num_blocks: Number of Transfor

Multi-level octree point cloud backed by Kaolin SPC. Stores per-level centroid xyz, mean rgb, and earliest frame index. The finest level (max

Orchestrates rendering, video encoding, and cleanup.

Offline Open3D renderer with octree LOD selection.

Manages resize operations for a scene. Lazily computes the ResizeTransform on first use and caches it. All resize operations are owned by thi

Builds a Scene from NPZ data, step by step. Each step stores intermediate state. Call .build() to produce the final immutable Scene.

ONNX-based sky segmenter.

CUDA-accelerated voxel grid using Morton coding + GPU sort.

3D旋转位置编码（3D RoPE）模块 核心思想：将RoPE扩展到3D空间（时间、高度、宽度），为视频或3D数据提供位置编码。 每个维度（t, h, w）独立使用RoPE，然后拼接起来。 公式： 对于3D位置 (f, h, w)（帧、高度

Evaluates camera pose accuracy using AUC metrics.

Camera state for rendering.

Causal self-attention module with KV cache support for streaming inference. Used by CasualBlockCamera in camera_head.py.

Handles client-specific UI and rendering logic.

Evaluates depth prediction quality using standard metrics. Computes metrics like absolute relative error, RMSE, and delta accuracies with sup

Orchestrates all evaluators for a single (gt_artifact, pred_artifact) pair. All evaluators share the interface: evaluate(gt_loader, pred_loader,

Feature fusion block.

Feature fusion block.

FlashInfer variant of the GCT attention layer. Uses FlashInferKVCacheManager for paged KV cache storage and FlashInfer attention kernels

Global viewer configuration.

Auto-calculate batch sizes to stay within GPU memory budget.

Fisheye undistortion (OPENCV_FISHEYE / equidistant, balance=0.0).

Manages keyframes and produces interpolated camera paths.

Scans Layer 1/2/3 eval outputs and prepares static report payloads.

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

Wraps dataset.evaluate_pointcloud() as a uniform evaluator.

Manages per-frame point cloud data and provides filtered subsets.

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

Generates a self-contained static report from Layer 1/2/3 eval outputs.

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

SDPA variant for streaming inference. Uses F.scaled_dot_product_attention with dict-based KV cache. No FlashInfer dependency required — w

Immutable scene data: point cloud + octree + metadata. The Scene does NOT know about cameras or perform culling. Rendering goes through octre

Manages persistent scene data caching (trajectory, point clouds) for faster loading.

Handles for a single frame in the scene.

Evaluates camera trajectories using ATE and RPE metrics.

Represents a single dataset/scene/method combination for viewing.

Holds all mutable viewer state.

Base class for all Aggregator implementations. Handles shared components: - Patch embedding (DINOv2 or conv) - Special tokens (camer

Abstract base class for all datasets. This class defines the interface that all dataset implementations must follow. Datasets are responsible

Abstract base class for all methods. This class defines the interface that all method implementations must follow. Methods are responsible fo

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

Loads classes dynamically without requiring registration.

DROID-W dataset loader (RGB + C2W TUM trajectory).

ETH3D multi-view stereo dataset (pi3 / custom_undistorted variant).

FlashInfer variant of causal block for GCT. Uses FlashInferAttention (FlashInfer paged KV cache + attention kernels). Supports optimized

Base class for GCT model implementations. Handles shared components: - Prediction heads (camera, depth, point) - Forward pass struct

Streaming GCT model with KV cache for efficient online inference. Features: - AggregatorStream with KV cache support (FlashInfer backend

Streaming GCT model with KV cache for efficient online inference. Features: - AggregatorStream with KV cache support (FlashInfer backend

General-purpose dataset loader for a single image folder or video file.

KITTI Odometry dataset loader (color, image_2).

LingbotMap model adapter for benchmark evaluation. Supports streaming and windowed inference modes via the upstream ``GCTStream`` model

Neural RGB-D dataset loader. Supports all 9 indoor scenes: breakfast_room, complete_kitchen, green_room, grey_white_room, kitchen, mo

Base class for render overlays.

Oxford Spires large-scale outdoor dataset.

SDPA variant for streaming inference. Uses F.scaled_dot_product_attention with dict-based KV cache. No FlashInfer dependency required.

7Scenes dataset loader.

Tanks and Temples dataset loader.

TUM RGB-D dataset loader (trajectory evaluation).

VBR dataset loader (RGB + C2W TUM trajectory + 3x3 intrinsics).