Leverage TSDF to post-process estimated depths for each pixel, find zero level of TSDF along ray (or closest to 0)

Align the origin (0,0,1) with the target point (x,y,1) in projective space. Method: rotate z to put target on (x'+,0,1), then rotate on Y to get

Saves the full training state, allowing exact resume: - model, optimizers, schedulers - discriminator (if used) - Ear

Compute 'overlapping' metrics based on a distance threshold between two point clouds.

Given an image and its intrinsics, we want to replicate a rectangular crop with an homography, so that the principal point of the n

Args: - depthmap (HxW array): - camera_intrinsics: a 3x3 matrix Returns: pointmap of absolute coordinates (HxWx3 arra

Convert Rotation matrix to quaternion, then calculate the location errors. original from PoseNet Paper

returns 3 values: 1 - reciprocal_in_P2: a boolean array of size P2.shape[0], a "True" value indicates a match 2 - nn2_in_P1: a int array

freeze bn layer by not require grad but still behave differently when model.train() vs. model.eval()

embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D)

grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or [n_cls_token+grid_size*grid_size, embe

build a prediction head for the decoder

returns (is_good, err_msg)

Check if a viewpoint is valid and overlaps significantly with a reference one.

open and convert all images in a list or folder to proper input format for DUSt3R

build a prediction head for the decoder

Normalize a vector.