Evaluate point cloud reconstruction for 7Scenes. Alignment uses align_gt_loader (gt resized to pred resolution) paired with pred_load

Evaluate 3D reconstruction against Neural RGB-D GT depth. Pipeline: 1. GT resized to pred resolution for pixel-level Umeyama coarse

Evaluate 3D reconstruction against ETH3D GT depth (7scenes-style pipeline). Pipeline: 1. GT resized to pred resolution -> pixel-lev

Extract video frames into a folder and return saved frame paths.

Convert camera extrinsics and intrinsics to a compact pose encoding. This function transforms camera parameters into a unified pose encoding form

Output (W, H) after transform, or None if not yet computed.

Forward pass. Args: images: Input images [B, S, 3, H, W] in range [0, 1] selected_idx: Which block indices t

Applies 2D rotary position embeddings to input tokens. Args: tokens: Input tensor of shape (batch_size, n_heads, n_tokens, dim).

前向传播：为3D输入（视频帧+patch）生成旋转位置编码 参数： - ppf (int): 帧数（patches per frame），当f_end为None时使用 - pph (int): 每帧的patch高度数

Forward pass with FlashInfer paged KV cache and attention. Args: x: Input tensor [B, N, C] kv_cache: FlashIn

Forward pass through the DPT head, supports processing by chunking frames. Args: aggregated_tokens_list (List[Tensor]): L

Forward pass. Args: x (tensor): input Returns: tensor: output

Forward pass. Returns: tensor: output

Forward pass. Returns: tensor: output

Forward pass to predict camera parameters. Args: aggregated_tokens_list (list): List of token tensors from the network;

Forward pass to predict camera parameters. Args: aggregated_tokens_list (list): List of token tensors from the network;

Deserialize from a dict (e.g., loaded from JSON). Args: d: Dictionary produced by to_dict() Returns: ResizeC

Create from a named preset, with optional style overrides.

GPU frustum cull. Returns int64 index array (K,) on CPU.

Generate pseudo intrinsics from image size.

Get camera intrinsic matrix from FOV and image size.

Get current camera state from client.

Get global evaluation defaults from base config.

Get list of frame IDs for a specific scene. Args: scene: Scene identifier Returns: List of frame IDs (intege

Get all frame identifiers for sequence. Args: scene: Sequence name (format: 'chess/seq-01') Returns: List of

Return sequential frame indices [0, 1, ..., N-1] for the sequence.

Return sequential frame indices [0 .. N-1].

Return frame indices for a scene. Indices are 0-based integers matching the img{N}.png / depth{N}.png file naming convention.

Get sorted list of frame IDs (parsed from JPG filenames). TAT images are named 000001.jpg, 000002.jpg, etc. (1-indexed). Args:

Return sequential frame indices [0, 1, ..., N-1].

Return 0-based frame IDs (one per pose line).

Get information about current KV cache state. Returns: Dictionary with cache statistics: - num_cached_bl

Get information about current KV cache state. Returns: Dictionary with cache statistics: - num_cached_bl

Get information about current KV cache state. Returns: Dictionary with cache statistics: - num_cached_bl

Return Path to ONNX model. Raises FileNotFoundError if not present. Args: model_path: Optional explicit path to ONNX model file.

Get raw value from base config.

Return the 0-based index of scene in the sorted scene list, or -1.

Get list of scene identifiers in the dataset. Returns: List of scene names/identifiers Note: Scene identifie

Get all sequence names (per sequence, not per scene). Returns: List of sequence identifiers in format 'scene_name/seq-XX'

Return list of sequence IDs as scene identifiers.

Return a single scene derived from the input name.

Return sorted list of scene names. Only directories containing a valid poses.txt are included, which correctly excludes any archives

Return TAT scenes whose directories exist under raw_data_root.

Discover all scene directories or return the configured whitelist.

Return all scene directory names that have images/ and poses_c2w.txt.

Check if a frame key was saved (from .complete.json frame_keys).

Check if a global key was saved (from .complete.json global_keys).

Check whether a frame-level data key is present in this directory.

Check if base config has a key.

Return True if both GT and method traj.txt files exist.

Streaming inference: process scale frames first, then frame-by-frame. This method enables efficient online inference by: 1.

Streaming inference: process scale frames first, then frame-by-frame. This method enables efficient online inference by: 1.

Windowed inference with keyframe detection and cross-window alignment. Each window is processed independently with a fresh KV cache.

ViT weight initialization, original timm impl (for reproducibility)

Load and preprocess images by center padding to square and resizing to target size. Also returns the position information of original pixels

Load per-frame custom data via context's __load_{key}_file__ method.

Load global-level custom data via context's __load_{key}_file__ method.

Load data for a single frame. Args: scene: Scene identifier frame_id: Frame ID (position in the post-sampling frame l

Load single frame data. Args: scene: Sequence name (format: 'chess/seq-01') frame_id: Frame number (integer)