[Task] image-feature-extraction (point matching) model support

Overview

Interest point detection and image pair matching models detect repeatable local features (keypoints) in images and match them across different views of the same scene. SuperPoint detects and describes keypoints; LightGlue efficiently matches descriptors between image pairs.

Agent Scenarios

• Visual localization agent: match a query photo to a map database for GPS-denied indoor navigation or AR anchoring
• 3D reconstruction agent: feed matched point pairs into structure-from-motion pipelines to build scene geometry from unordered photos
• Image stitching / panorama agent: align overlapping frames from drone or 360-camera feeds before compositing
• Inspection agent: compare before/after images of structures (bridges, PCBs) by matching keypoints to detect changes or defects

ModelKit Integration

Models must pass the full wmk pipeline on all EPs:

wmk config → wmk build (ONNX export) → wmk perf → wmk eval

Acceptance Criteria

• ETH-CVG/lightglue_superpoint
• magic-leap-community/superpoint

Start by locating the ModelKit integration entry points for ETH-CVG/lightglue_superpoint and magic-leap-community/superpoint. Run the models through wmk config, wmk build for ONNX export, wmk perf, and wmk eval; done means both models pass the full pipeline on all EPs.