cs.CVAug 20, 2026
SaveUnified and Efficient Point-Line Local Features
Organizations: Department of Computer Science, ETH Zurich, Switzerland · Microsoft Spatial AI Lab
Abstract
Multi-view computer vision pipelines typically rely on accurate sparse keypoints and robust descriptors. While incorporating line features has shown clear benefits for matching and pose estimation, existing point-line approaches remain inefficient: they detect points and lines separately, use increasingly heavy networks, and depend on CPU-bound heuristics that hinder real-time performance. We introduce a Unified Efficient Points and Lines (UPAL) feature extractor that jointly extracts keypoints, line segments, and feature descriptors within a single lightweight architecture. A shared backbone provides common representations that feed different branches for point and line features. Line segments are recovered through an accelerated post-processing stage, an enhanced and highly efficient variant of the LSD algorithm. UPAL matches or exceeds state-ofthe-art performance in both point and line applications while significantly reducing computational cost, achieving, for instance, a 4x speedup and 10x smaller memory footprint over the ALIKED + DeepLSD pipeline. Code is publicly available at https://github.com/francois141/upal.
Figures & tables
Figure 1: In this paper, we propose UPAL, a joint point-line feature extractor that outperforms most existing methods (here, on hybrid point-line visual localization with pose accuracy under 5cm / 5 ∘ ) while being substantially more lightweight . This is enabled by our compact architecture, joint point-line training with distilled supervision, and an efficient variant of the LSD algorithm. Results are shown on the Stairs scene of the 7Scenes dataset [ 61 ] with the setup described in Sec. 4.4 .
Figure 2: Overview of our network architecture. We use a shared convolutional encoder to generate a multi-scale feature map (orange). We then use three decoder branches: a keypoint detection branch (blue) with a Score Map Head (SMH) followed by Differentiable Keypoint Detection (DKD), a Sparse Deformable Descriptor Head (SDDH) to obtain descriptors (purple), and a branch predicting a distance field (green). The latter is then passed to our fast LSD to extract the line segments.
Figure 3: Visualization of the outputs of the network. The shared backbone receives an input image (a) and encodes it into bottleneck features (b). We observe that the predicted keypoint score map (c) and line distance field (d) show some similarity (e.g. keypoints are often line endpoints or intersections), justifying their joint prediction.
Figure 4: Line detection examples. In spite of its efficiency, UPAL outputs generic lines with quality similar to more expensive methods.
| Method | AUC @1px | AUC @3px | AUC @5px | Latency (ms) |
| Lightweight Methods | ||||
| SIFT [ 39 ] | 35.1 | 65.0 | 75.3 | 237 |
| SuperPoint [ 10 ] | 35.6 | 65.2 | 75.7 | 50 |
| DISK [ 65 ] | 29.9 | 61.8 | 73.2 | 100 |
| XFeat [ 47 ] | 34.4 | 62.2 | 73.3 | 44 |
| Wireframe [ 17 ] | 38.3 | 66.4 | 76.1 | 308 |
Table 1: Homography estimation on HPatches [ 4 ] . We report the latency in ms and the Area Under the Curve (AUC) of correctly predicted homographies for various pixel error thresholds. Methods are grouped into lightweight and heavyweight categories. Within each category, the best result is shown in bold , and the second best is underlined .
| Method | AUC @5 ∘ | AUC @10 ∘ | AUC @20 ∘ | Latency (ms) |
| Lightweight Methods | ||||
| SIFT [ 39 ] | 40.1 | 53.6 | 64.4 | 729 |
| SuperPoint [ 10 ] | 51.3 | 64.4 | 73.7 | 78 |
| DISK [ 65 ] | 53.9 | 66.1 | 75.3 | 335 |
| ALIKED [ 80 ] | 59.8 | 72.0 | 81.1 | 79 |
| XFeat [ 47 ] | 44.6 | 59.0 | 70.2 | 28 |
Table 2: Relative pose estimation on MegaDepth [ 34 ] . The latency and AUC of correctly retrieved poses under various angular error thresholds are reported. Methods are grouped into lightweight and heavyweight categories. Within each category, the best result is highlighted in bold and the second best is underlined . Results reported in parentheses are not considered for ranking due to overlap between the training data and the MegaDepth benchmark test set.
| Method | AUC @5 ∘ | AUC @10 ∘ | AUC @20 ∘ | Latency (ms) |
| Lightweight Methods | ||||
| SIFT [ 39 ] | 6.7 | 14.9 | 24.2 | 742 |
| SuperPoint [ 10 ] | 15.5 | 29.8 | 43.4 | 83 |
| DISK [ 65 ] | 10.1 | 19.2 | 29.5 | 348 |
| ALIKED [ 80 ] | 6.9 | 13.3 | 20.1 | 82 |
| XFeat [ 47 ] | 16.4 | 31.2 | 45.7 | 27 |
Table 3: Relative pose estimation on ScanNet [ 8 ] . The latency and AUC of correctly estimated poses under various angular error thresholds are reported. Methods are grouped into lightweight and heavyweight categories. Within each category, the best result is highlighted in bold and the second best is underlined .
| HPatches | RDNIM | |||||||||||||||||
| Loc Err | H Estim | Rep | Time | Loc Err | H Estim | Rep | Time | |||||||||||
| 50l | 300l | 1px | 3px | 5px | 1px | 3px | 5px | (ms) | 50l | 300l | 1px | 3px | 5px | 1px | 3px | 5px | (ms) | |
| LSD [ 68 ] | 0.50 | 1.42 | 58.3 | 88.3 | 90.9 | 26.4 | 53.6 | 61.6 | 150 | 1.78 | 1.85 | 7.4 | 44.8 | 57.6 | 9.2 | 30.0 | 37.7 | 85 |
| ELSED [ 63 ] | 0.71 | 1.63 | 40.7 | 65.8 | 67.9 | 18.8 | 41.8 | 50.8 | 67 | 1.92 | 1.96 | 5.9 | 30.8 | 40.9 | 8.9 | 25.9 | 30.9 | 40 |
| TP-LSD [ 27 ] | 1.49 | 1.61 | 10.9 | 34.6 | 42.6 | 11.6 | 42.9 | 57.2 | 65 | 1.93 | 1.93 | 0.2 | 6.3 | 14.2 | 4.9 | 33.3 | 47.5 | 52 |
| SOLD2 [ 45 ] | 1.09 | 1.29 | 23.7 | 48.5 | 55.4 | 19.8 | 44.9 | 51.5 | 745 | 1.39 | 1.41 | 0.6 | 9.4 | 16.0 | 8.8 | 20.0 | 24.0 | 546 |
Table 4: Line detection evaluation on HPatches [ 4 ] and RDNIM [ 43 ] . We compare the line localization error (Loc Err) on a fixed set of l lines, homography estimation (H Estim), and line repeatability (Rep) at various pixel thresholds. Horizontal lines separate Traditional / Learned / Joint methods.
| Method | T / R err. | Acc. | |
| Points | SuperPoint [ 10 ] | 0 6.2 / 1.65 | 36.6 |
| DISK [ 65 ] | 0 7.2 / 2.16 | 28.5 | |
| ALIKED [ 80 ] | 0 7.8 / 1.98 | 35.0 | |
| PLNet [ 70 ] | 0 6.2 / 1.68 | 38.2 | |
| UPAL (Ours) | 0 5.1 / 1.39 | 49.1 | |
| Points + Lines | SuperPoint [ 10 ] +DeepLSD [ 44 ] | 0 5.0 / 1.33 | 49.6 |
Table 5: Visual localization on Stairs scene of the 7Scenes dataset [ 61 ] . We report the median translation / rotation errors (cm / deg) and pose accuracy at 5 cm / 5° threshold. Please refer to our supplementary material for results on all scenes.
| Method | Completeness (%) | Accuracy (%) | |||||
| 1cm | 2cm | 5cm | 1cm | 2cm | 5cm | ||
| SuperPoint [ 10 ] | 0.17 | 0.81 | 4.34 | 49.52 | 64.73 | 80.01 | |
| DISK [ 65 ] | 0.33 | 1.18 | 4.41 | 69.22 | 81.10 | 91.29 | |
| ALIKED [ 80 ] | 0.13 | 0.56 | 2.76 | 56.61 | 70.78 | 85.18 | |
| Wireframe [ 17 ] | 0.19 | 0.98 | 5.41 | 50.98 | 66.44 | 82.50 | |
| UPAL (Ours) | 0.30 | 1.09 | 4.45 | 64.80 | 77.57 | 88.87 | |
Table 6: Evaluation on multi-view point triangulation on ETH3D [ 59 ] .
| Nb params | Latency (ms) | ||
| GPU | CPU | ||
| SuperPoint [ 10 ] + DeepLSD [ 44 ] | 9.8M | 293 | 2259 |
| ALIKED [ 80 ] + DeepLSD [ 44 ] | 9.2M | 286 | 2239 |
| ALIKED [ 80 ] + M-LSD [ 22 ] | 1.3M | 54 | 525 |
| ALIKED [ 80 ] + TP-LSD [ 27 ] | 25M | 56 | 1582 |
| DaD [ 14 ] + DeDoDe v2 [ 15 ] + ScaleLSD [ 29 ] | 120M | 147 | >200K |
Table 7: Runtime evaluation. Parameter counts are reported, and latencies are averaged over 500 sequential detections on an NVIDIA RTX 2080 Ti GPU and Intel Core i9-13900K CPU.
| Loc Error | H estimation | Repeatability | Latency (ms) | ||
| (1) | DeepLSD [ 44 ] | 1.432 | 90.6 | 27.2 | 233 |
| (2) | DeepLSD no AF | 1.341 | 90.6 | 28.2 | 218 |
| (3) | LSD [ 68 ] | 1.414 | 91.1 | 26.3 | 36 |
| (4) | (3) + parall. ang. comp. | 1.414 | 91.1 | 26.3 | 27 |
| (5) | (4) + stride 2 | 1.455 | 90.6 | 26.3 | 19 |
| (6) | (5) + subset | 1.476 | 90.6 | 26.3 | 17 |
Table 8: Ablation study on the HPatches dataset [ 4 ] . We compare the localization error, homography estimation, repeatability, and latency for variants of our approach.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | AUC@5 ∘ | AUC@10 ∘ | AUC@20 ∘ |
| Ours (points only) | 55.2 | 67.6 | 76.4 |
| Ours (joint training) | 58.2 | 70.8 | 79.4 |
| Improvement | +3.0 | +3.2 | +3.0 |
Table 9: Impact of joint point and line training on MegaDepth [ 34 ] relative pose estimation.
| Nb params | Latency | |
| SuperPoint [ 10 ] + DeepLSD [ 44 ] | 707 | |
| ALIKED [ 80 ] + DeepLSD [ 44 ] | 716 | |
| DaD [ 14 ] + DeDoDe v2 [ 15 ] + ScaleLSD [ 29 ] | 3474 | |
| UPAL (Ours) | 186 |
Table 10: Runtime evaluation. Parameter counts and latencies (ms) are averaged over 500 sequential detections on an NVIDIA GeForce GTX 1050 Ti with 4 GB of VRAM. UPAL runs in 186 ms, about 18 faster than DaD + ScaleLSD.
Figure 5: Line detection examples. Unlike wireframe-based methods [ 29 , 17 ] that focus solely on structural lines, UPAL produces more general and repeatable line detections that better support downstream tasks, while remaining significantly more lightweight. The last four images are challenging examples for UPAL and other detectors.
Figure 6: Point matching on HPatches (yellow correct, red wrong with respect to the ground-truth homography; only the most salient matches are shown). UPAL generally matches keypoints reliably under both viewpoint and illumination changes. On repetitive structures the salient corners still match, with errors confined to weaker detections, whereas specular surfaces and especially self-similar texture under large viewpoint changes are harder and yield many incorrect matches.
Figure 7: Line matching on HPatches (green correct, red wrong with respect to the ground-truth homography; only the longest matched segments are shown). UPAL matches lines well under illumination changes on structured man-made scenes, but its line branch is not viewpoint-robust: the same viewpoint pairs that are matched reliably by points are almost entirely mismatched as lines.
More in this topic
Local Feature Matching
Local feature matching finds corresponding points or regions across images to support geometric estimation, localization, reconstruction, and recognition.
See all 34 papers