Device path: run inference, D2H the scores only, select keypoints on the host, then gather their descriptors on the GPU.

Batched stereo device path: preprocess L and R into one {2,1,H,W} input, a single enqueue, then select and gather each batch slice.

KITTI t_rel (fraction) and r_rel (deg/m) averaged over all segment lengths. gt_poses / est_poses are [N,4,4] homogeneous camera-to-world poses.

TartanAir/TartanGround pose file: per-frame 'tx ty tz qx qy qz qw' (no timestamp).

Return the number of accepted loop closures so far.

LightGlue keypoint normalization: (kpt - size/2) / (max(w,h)/2).

Thread-management hooks: plain flag toggles.

Relative Pose Error over fixed deltas (drift).

Shared keypoint-select (host) and on-device gather for one image or batch slice. `scores_host` and `grid_device` point at this slice's score map or FP

Associate two timestamped trajectories by nearest timestamp.