Return the dense score heatmap and descriptor grid for ONNX export.

Thin wrapper exposing tensor I/O (no dict, no extractor) for ONNX export.

SuperPoint detector and descriptor (DeTone, Malisiewicz, Rabinovich, CVPRW 2019). The conv1a..conv4b shared encoder, convPa/convPb detector head,

Extractor that alternates left/right keypoints on successive calls. The front-end calls extract() twice per frame (left, then right); the right set is

Source-agnostic cosine-similarity index over L2-normalized global descriptors. Both the real EigenPlaces recognizer and the unit-test stub delegate ad

Hold a fixed pool of N device slots of FP16 descriptor rows. Size each slot for the worst case (max_keypoints * dim half-floats); DeviceDescriptors::c

Hold a device pointer and shape for descriptors resident in a DescriptorPool slot. Opaque to the GPU-free core (no nvinfer or CUDA types here); the in

EigenPlaces (ResNet18) global-descriptor place recognizer on TensorRT. Produce one L2-normalized descriptor per image; delegate retrieval to CosineDes

Hold backend-agnostic feature-extractor output. Per-keypoint score lives in cv::KeyPoint::response.

Motion-only pose solve for an ordinary frame against the last keyframe's points.

Manage a fixed free-list over N slot indices. CUDA-free slot bookkeeping, unit-testable without a device, separate from the device allocation that Des

Tier-2 backend: a batch pose graph over keyframe poses. Nodes are 6-DOF Pose3 keyframes (symbol X(keyframe_id)); edges are BetweenFactor<Pose3> odomet

Pluggable feature extractor.

Pluggable feature matcher. Two descriptor sources: host cv::Mat (loop closure, descriptors in the host keyframe DB) and DeviceDescriptors (live tracki

Pluggable place recognition. The descriptor source (a learned global-descriptor model) lives behind this interface; CosineDescriptorIndex holds the sh

Identity index matcher: left i <-> right i.

Identity index matcher: left i <-> right i up to min(N, M).

Identity matcher: returns i->i for the first `n` features. Used for both per-frame tracking and loop verification (sync mode -> no concurrency, safe t

Insertion-ordered store of keyframe records with O(1) id lookup.

One keyframe's persisted state for loop closure. The authoritative optimized pose lives in the GlobalPoseGraph (keyed by keyframe_id); pose_at_insert

LightGlue matcher (TensorRT). Implement superslam::IFeatureMatcher. Expect this ONNX I/O (matches utils/convert_lightglue_to_onnx.py): inputs : kpts0

Hold a retrieved place-recognition match: a database keyframe and its similarity score.

Pose-graph loop closure with implicit landmarks. Retrieval (place recognition) finds a revisited keyframe; LightGlue and the existing FrameTracker geo

Tunables (env-overridable, matching the repo's SUPERSLAM_* convention).

Outcome of a loop-closure attempt for one query keyframe.

Hold the result of matching two feature sets: query and train index pairs (cv::DMatch) and, optionally, a raw score matrix. Keep in the global namespa

Matcher that returns identity correspondences i->i for the first `n` features.

Pose-only monocular reprojection factor (2 residuals: u, v).

Pose-only stereo reprojection factor (3 residuals: uL, uR, v).

Stream the live trajectory, stereo cameras, on-demand point cloud, text logs, and plots to a Rerun viewer. Spawn the interactive viewer, or record to

Convert an RGB-D image and depth map to a StereoFrame. Depth gives the right-image coordinate directly (uR = uL - bf/Z); the front end needs no featur

Sparse point-cloud map for visualization and export. Hold each keyframe's depth-valid feature points in the keyframe camera frame; cloud() lifts them

Lean, GTSAM-native per-frame data. Pose is Twc (camera-in-world).

Convert a stereo image pair to a StereoFrame. Hold backend-agnostic inference interfaces (no knowledge of TensorRT). Leave pose at identity for the es

Identity place recognizer: descriptor == image; retrieval via the shared cosine index.

Place recognizer whose descriptor IS the passed image row (so the test controls retrieval exactly), backed by the shared cosine index.

Public facade: the one object users construct. Own the SuperPoint and LightGlue front-end and the GTSAM stereo VO; everything internal is Twc.

Debounce loops: accept a place match only after K consecutive queries agree on the same locale. Defeat one-off perceptual-aliasing false positives. Sa

Orchestrate sliding-window stereo VO, optionally upgraded to SLAM with pose-graph loop closure: StereoFrame to per-frame tracker to keyframe gate to b

Hold a bounded sliding window of the last K keyframe poses, optimized as a batch with smart stereo factors (landmarks marginalized). No persisted map.