Modulate the input tensor using scaling and shifting parameters.

Convert pose encoding to extrinsics (c2w) and move to CPU.

Fused pre-attention ops for single-frame streaming (Phase 2). Computes q/k/v from x, applies q_norm/k_norm/RoPE, and converts to [tpf

Apply sky masking (if configured).

Process a scene and return results. Args: gt_artifact: BSSArtifact for the ground truth directory Returns: D

Render NPZ predictions to video using the rgbd_render pipeline. Args: npz_path: Path to NPZ file or per-frame NPZ directory. outp

Convert rotation matrix or batch of matrices to quaternion(s). Uses Shepperd's method for numerical stability. Args: R: (3, 3) or (.

Run ONNX sky segmentation on a BGR image and return an 8-bit score map.

Run ONNX sky segmentation on a single BGR image and return an 8-bit score map.

Save depth map as pseudo-color visualization. Args: depth: HxW depth map (meters) output_file: Output JPEG file path min_

Segment sky from an image using ONNX model. Args: image_path: Path to the input image skyseg_session: ONNX runtime inference

Segment sky from an image array using ONNX model. Args: image: Input image as numpy array (H, W, 3) or (3, H, W), values in [0, 1] o

Batch version of segment_sky_from_array.

Draw frame counter text on rendered image. Modifies in-place.

Return images downsampled so the long edge = long_edge pixels.

Save config to YAML file.

Applies a 4x4 transformation to a set of points. Args: transformation: Transformation matrix points: Points to be transforme

Unproject a batch of depth maps to 3D world coordinates. Args: depth_map (np.ndarray): Batch of depth maps of shape (S, H, W, 1) or

Update point size and background.

Update point cloud based on current GUI selections.

Validate method output format. Args: output: Output dictionary from process_scene() num_frames: Expected number o

Unproject depth and build octree.

Bring all windows into the first window's coordinate frame, then stitch. Iterates over consecutive window pairs, estimates the pairwise

Bring all windows into the first window's coordinate frame, then stitch. Iterates over consecutive window pairs, estimates the pairwise

Solve gt = s * pred + t using least squares.

Apply sliding window eviction to KV cache.

Apply sliding window eviction to KV cache BEFORE attention. This ensures current frame only attends to frames within the sliding win

Apply random sampling. Args: frames: Original frame list config: Random sampling configuration Returns:

Apply a ResizeTransform to an image. Pipeline: 1. Resize to intermediate_size 2. Crop [content_y0:content_y0+content_h, content_x0:co

Apply sequence (stride or uniform) sampling. Args: frames: Original frame list config: Sequence sampling configuratio

Apply sky segmentation mask to confidence scores.

NumPy port of visibility_kernel2. For each non-zero pixel v at (row, col) with depth d: - Back-project to 3D: v = ((col-cx)*d/fx, (ro

Thin wrapper around the compiled visibility2 CUDA kernel.

Build frame_blocks and global_blocks. Subclasses implement mode-specific block creation. Must create: - self.frame_

Generate all pairwise indices for N frames.

Build patch embedding layer. Supports: - "conv": Simple convolutional patch embedding - "dinov2_*": DINOv2 ViT varia

Build a tube mesh following camera trajectory with per-segment color. Args: positions: (N, 3) camera positions. radiu

Infer the Freiburg camera id (freiburg1/2/3) from a scene directory name.

Check if COLMAP binary exists on PATH.

Clear cache for current data source and reset the displayed scene.

Collect sorted image paths from the image directory.

Unproject all frames and return (all_xyz, all_rgb, all_frames) on CPU.

Convert COLMAP camera parameters to [fx, fy, cx, cy]. Supports: SIMPLE_PINHOLE, PINHOLE, SIMPLE_RADIAL, RADIAL, OPENCV.

Compute base point size from average pixel spacing using first frame. Args: depth_list: List of depth maps intrin

Compute a single bird's-eye camera from trajectory + point cloud.

Compute centroid xyz / mean rgb / min frame at every level.

Computes frequency components for rotary embeddings. Args: dim: Feature dimension (must be even). seq_len: Maximum se

Compute depth metrics on concatenated valid pixels.

Compute a ResizeTransform for the given mode. Args: orig_w: Original image width orig_h: Original image height

Copy current camera view parameters to global clipboard.

Cover-fit resize + center-crop to ``self.target_size``. Picks the smallest scale s such that the resized image still covers the targe

Cover-fit resize + center-crop to ``self.target_size``. Picks the smallest scale s such that the resized image still covers the targe

Create trajectory path with rainbow gradient colors (segmented lines).

Estimate per-batch scale as the median depth ratio anchor/target.

Estimate per-batch scale as the median depth ratio anchor/target.

Detect method output image size from the first non-None RGB frame. Args: frame_data_list: List of frame data dicts Retur

Discover __save_{key}_file__ methods on context.

Discover and filter scenes from input_folder or video, returning (scenes, video_images).

Apply Eye-Dome Lighting post-processing. For each foreground pixel, compares log2(depth) with 8 neighbours at the given *radius*. Large dept

Apply Eye-Dome Lighting post-processing for depth cues.

Embed images and prepare for attention processing. Handles: - Image normalization - Patch embedding - Specia