[lukapercic]

Motivation

• Provide a custom extractor that intercepts SuperPoint detection to supply EdgeDrawing endpoints as keypoints for description.
• Allow using pure SuperPoint, pure EdgeDrawing, or a combined mode so descriptors can be sampled at ED endpoints or merged with standard detections.
• Enable describing specific pixels (line endpoints) using SuperPoint's dense descriptor maps to support edge-aware features.

Description

• Add custom_superpoint.py implementing SuperPointWithED, a subclass of SuperPoint that accepts config and ed_config.
• Implement run_edgedrawing which runs OpenCV EdgeDrawing (cv2.ximgproc) on a grayscale image and returns unique line endpoints as (N, 2) points.
• Override forward to obtain pred = self.net(image), extract standard SuperPoint keypoints (with NMS and top-k filtering), obtain ED keypoints, merge them according to ed_conf['mode'] (superpoint, edgedrawing, combined), and sample descriptors and scores with sample_descriptors.
• Handle RGB->grayscale conversion, batch-size-1 simplification for ED, and safe fallbacks for empty keypoint sets.

Testing

• No automated tests were executed for this change.
• The new file custom_superpoint.py was added and committed without running test suites.
• Manual import/run was not performed in automated CI in this rollout.
• Further unit or integration tests should be added to validate descriptor sampling and ED integration across batches.

---

Codex Task

[lukapercic]

still testing