UPAL¶

Development line: project:upal · thread upal-development
Last event: 2026-08-24 · 1 dated since 2026-08-24 · Researched: 2026-09-05 · confidence: medium

What it is¶

UPAL is a standalone PyTorch feature extractor for CV and SLAM practitioners. It combines point detection, line detection, and local descriptors in one network.

• sub-pixel keypoints and confidence scores;
• 128-dimensional L2-normalized descriptors;
• keypoint heatmaps and line-distance fields, with bundled points_lsd post-processing for segments.

Development line¶

• 2026-08-24 — UPAL GitHub repository referenced. The repository appeared publicly, but no first-party commit, tag, or release ties to this date, so it does not establish a new version.

What changed¶

2026-08-20: arXiv v1 introduced UPAL as a shared-backbone point-line feature extractor with public code. It reported roughly 4× speedup and a 10× smaller memory footprint than ALIKED + DeepLSD. 2026-08-24: UPAL's repository appeared publicly, but no first-party commit, tag, or release ties to this date, so it does not establish a new version.

How to use this¶

As of 2026-08-24, start from the UPAL GitHub repository for project research, because the evidence supports no narrower operational guidance.

1. Check out the repository, make a Python 3.10+ environment, and install the package with pip install -e .. — https://github.com/francois141/upal
2. Initialize recursive submodules and install third_party/points_lsd. For CUDA, install the matching PyTorch build before building that extension. — https://github.com/francois141/upal
3. Run demo_inference.py, demo_match_points.py, and demo_match_lines.py to test the supported inference, point-matching, and line-matching paths. — https://github.com/francois141/upal
4. To use UPAL in an application, load weights/upal.tar, pass a normalized [0,1] tensor shaped B×1×H×W or B×3×H×W, and read out keypoints, descriptors, heatmaps, and the line-distance field. — https://github.com/francois141/upal

Best practices¶

• Install a CUDA-compatible PyTorch build before compiling points_lsd, because the line path depends on that native submodule. — https://github.com/francois141/upal
• Test all three bundled demos before integration, because inference, point matching, and line matching are separate supported paths. — https://github.com/francois141/upal
• Do not generalize the speed figure across deployments: the published 70 ms result used 500 sequential 800×800 images, batch size one, and an RTX 2080 Ti. — https://arxiv.org/html/2608.19894v1

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• We found no primary-source commit, tag, or release dated 2026-08-24, so we cannot treat that date as a version release.
• The repository documents standalone inference and demos, not a supported production service, training workflow, or operating-system matrix.
• We found no verified community installation report, so target-runtime builds and performance remain untested.

Sources¶

Agent brief¶

• Subject: project:upal, thread upal-development, 1 dated events 2026-08-24 → 2026-08-24.
• Practical note: As of 2026-08-24, start from the UPAL GitHub repository for project research, because the evidence supports no narrower operational guidance.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.