Return a new BSSLoader with the same artifact, resize_context, and context.

Forward pass of the GCT model. Args: images: Input images [S, 3, H, W] or [B, S, 3, H, W], in range [0, 1] q

Apply 3D rotary position embedding using real arithmetic (torch.compile-safe). Equivalent to complex multiplication but avoids torch.view_as_comp

Clean KV cache in aggregator. Call this method when starting a new video sequence to clear cached key-value pairs from previ

Get the number of frames from .complete.json metadata. Returns: Number of frames in this scene Raises: Value

Load cached scene data. Returns: Tuple of (trajectory_dict, point_clouds_list, is_aligned)

Override cfg field from args. - Without --config: always applies the CLI value (preserves legacy behavior, including baked-in argpa

Return the BSSArtifact for a GT or method directory. Args: dataset_name: Dataset config name scene: Scene name

Set _skip_append flag on all KV caches (aggregator + camera head). When skip=True, attention layers will attend to [cached_kv + current_kv]

Load camera trajectory. Returns: np.ndarray of shape (N, 4, 4), with NaN 4x4 matrices for missing poses. N is always

Load colored point cloud in world coordinates. This method loads colored 3D point clouds in world coordinate system from two sources:

Compute the inverse of each 4x4 (or 3x4) SE3 matrix in a batch. If `R` and `T` are provided, they must correspond to the rotation and transl

Load from JSON.

Convert a pose encoding back to camera extrinsics and intrinsics. This function performs the inverse operation of extri_intri_to_pose_encoding,

Quaternion Order: XYZW or say ijkr, scalar-last Convert rotations given as quaternions to rotation matrices. Args: quaternions:

Run the viewer.

Get the original GT frame indices for stored frames. For sparse SLAM outputs (K < N), returns the 'frame_index_map' list stored in .c

Load all RGB images in frame index order (000000.png, 000001.png, ...). For sparse outputs (e.g. SLAM keyframe-only), frames without a saved

Create dataclass instance from dict, ignoring unknown keys.

Reset the viewer camera to look at scene center from a given direction. Args: direction: Unit vector pointing FROM the scene cent

Toggle between config mode (True) and viewing mode (False). Config mode: Load enabled, rebuild params enabled, dropdowns enabled, Clear Cach

Append one frame's K/V tensors to the two-stream cache. Token layout must be: [camera, reg0, ..., regN, scale, patch0, ..., patch

Apply a 4x4 transformation matrix to a set of 3D points. This function transforms 3D points using a homogeneous transformation matrix. Only t

Compute cross-frame attention using FlashInfer BatchPrefillWithPagedKVCacheWrapper. When self.force_fp32 is True, gathers all vis

Compute nearest neighbor distances from source points to target points. For each point in the source point cloud, finds the distance to its neare

optical flow induced by camera motion

Load all depth maps in frame index order. Returns: List of HxW depth maps (float32, meters), resized if resize_context is set.

Undo the most recent ``append_frame()`` for *block_idx*. This reverses all three sub-operations of ``append_frame``: patch page al

Save scene data to cache. Args: traj_dict: Dictionary mapping frame index (int) to C2W matrices point_clouds: List

Helper function to slice, expand and flatten tokens. Args: token: Token tensor [1, 2, N, C] where first index is for first frames

Convert valid pose array to evo PoseTrajectory3D. Orthogonalizes rotations at the evo boundary to satisfy SO(3) validation. Args: po

Resize spatial data to GT resolution if it differs from method resolution. Args: data: Input array (HxW or HxWxC)

Execute a COLMAP sub-command; return True on success.

Drop the leading batch dimension for single-sequence demo outputs.

Drop the leading batch dimension for single-sequence demo outputs.

Interpolate keyframes to get camera at any frame.

Convert rotations given as rotation matrices to quaternions. Args: matrix: Rotation matrices as tensor of shape (..., 3, 3). Re

Reset all per-block state for a new sequence.

Full pipeline: render → encode → cleanup.

Save dataset-level aggregation entry.

Rebuild point cloud geometry using octree LOD selection.

Align all camera poses to the first camera's coordinate frame.

Drop frame pairs where GT or pred pose contains NaN/Inf. Logs a WARNING if any pairs are dropped (indicates some GT frames lack valid poses,

Create a ramp that fades from a base color to darker.

Set defer-eviction flag on FlashInfer manager and SDPA caches. While True, eviction is suppressed so that rollback can cleanly undo t

Set defer-eviction flag on FlashInfer manager and SDPA caches. While True, eviction is suppressed so that rollback can cleanly undo t

Set _skip_append flag on all KV caches (aggregator + camera head). When skip=True, attention layers will attend to [cached_kv + current_kv]

Set _skip_append flag on all KV caches (aggregator + camera head). When skip=True, attention layers will attend to [cached_kv + current_kv]

Update frame visibility based on timeline.

Clean KV cache in aggregator. Call this method when starting a new video sequence to clear cached key-value pairs from previ

Clean KV cache in aggregator. Call this method when starting a new video sequence to clear cached key-value pairs from previ

Evict old window patch pages (recycle to free list). Special pages are NEVER evicted. Scale pages are NEVER evicted.

Get image dimensions from metadata. Returns: (width, height) tuple

Get workspace path from configuration.

Check if evaluation result exists for given type.

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.

List all scenes in a dataset directory (sanitized, flat format).

A quick start function to load and preprocess images for model input. This assumes the images should have the same shape for easier batching,