extract confidence from prediction head output

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

Create configs and perform basic setups.

Apply undistortion to the normalized tracks using the given distortion parameters once. Args: params (torch.Tensor or numpy.ndarray)

Processes specialized tokens with shape (1, 2, X, C) for multi-frame processing: 1) Uses the first position (index=0) for the first frame onl

Predict in the correct aspect-ratio, then transpose the result in landscape and stack everything back together.

Init. Args: scale_factor (float): scaling mode (str): interpolation mode

Init. Args: scale_factor (float): scaling mode (str): interpolation mode

Init. Args: scale_factor (float): scaling mode (str): interpolation mode

Simple UV positional embedding added to feature maps.

Simple UV positional embedding added to feature maps.

Accept (4,4) or (3,4) extrinsic parameters, return (4,4) homogeneous matrix.

Accept (4,4) or (3,4) extrinsic parameters, return (4,4) homogeneous matrix.

Backprojection operation: from image depths to 3D points

return_all_blocks: if True, return the features at the end of every block instead of just the features from the l

Core forward implementation for DPT head. Args: token_list: List of transformer tokens from each layer, [B, S, N, C]

Args: aggregated_tokens_list (List[Tensor]): List of token tensors from different transformer layers. images (Tensor)

Implementation of the forward pass through the DPT head. This method processes a specific chunk of frames from the sequence.

Implementation of the forward pass through the DPT head. This method processes a specific chunk of frames from the sequence.

Implementation of the forward pass through the DPT head. This method processes a specific chunk of frames from the sequence.

Implementation of the forward pass through the DPT head. This method processes a specific chunk of frames from the sequence.

Implementation of the forward pass through the DPT head. This method processes a specific chunk of frames from the sequence.

Implementation of the forward pass through the DPT head. This method processes a specific chunk of frames from the sequence.

Implementation of the forward pass through the DPT head. This method processes a specific chunk of frames from the sequence.

Implementation of the forward pass through the DPT head. This method processes a specific chunk of frames from the sequence.

Join URL parts safely.

Infer per-frame KV token count from a cache produced for a known number of frames.

Prepare transforms for PRoPE-style positional encoding.

Prepare transforms for PRoPE-style positional encoding.

Process frame attention blocks. We keep tokens in shape (B*S, P, C).

Process global attention blocks. We keep tokens in shape (B, S*P, C).

Process global attention blocks. We keep tokens in shape (B, S*P, C).

Process global attention blocks. We keep tokens in shape (B, S*P, C).

Process the next task in the queue.

Map raw relative confidence logits to strictly positive weights.

Resolve translation/rotation confidences and whether true split weighting is active.

Mark a fresh dynamic segment using replay-local frame IDs only.

Build pairwise relative pose encoding for valid sparse edges from absolute pose encoding.

Stack per-frame tensors along the view dimension for batched metric inference.

Build a normalized summary vector for novelty and similarity checks.

Update visualized cameras based on current source (pcd0 vs pcd).

Return whether online inference should route global attention through FlashInfer pages.

Accumulate adjacent SIM(3) transforms into transforms from frame 0 to each frame. Args: transforms: list of (s, R, t) tuples Returns