A computer-vision preprint describes UPAL as a single-network system that jointly predicts image points, line segments and the feature descriptors used to match them. The design targets lower inference time and memory use than running separate point and line detectors.
Across 500 sequential detections, UPAL used 0.78 million parameters and recorded mean latencies of 70 milliseconds on a GPU and 976 milliseconds on a CPU. The paper’s conclusion reports a fourfold speedup over combining standalone components and a memory footprint at least one order of magnitude smaller. On an NVIDIA GeForce GTX 1050 Ti with 4 GB of VRAM, it ran in 186 milliseconds, reported as about 18 times faster than DaD plus ScaleLSD.
Benchmark results
On HPatches, a point benchmark built from 580 image pairs with illumination or viewpoint changes, UPAL reported homography-estimation AUC values — a summary score shown at different error cutoffs — of 39.2% at 1 pixel, 67.9% at 3 pixels and 77.7% at 5 pixels. Its reported latency was 62 milliseconds.
For relative camera-pose estimation, the reported AUC on 1,500 MegaDepth image pairs was 58.2% at 5 degrees, 70.8% at 10 degrees and 79.4% at 20 degrees, with 80 milliseconds of latency. On 1,500 ScanNet pairs, the corresponding figures were 15.6%, 28.9% and 41.4%, with 83 milliseconds of latency.
The line tests included 1,722 RDNIM image pairs. On HPatches, UPAL reported a localization error of 1.42 at the 300-line setting, homography estimation of 90.9% and repeatability of 61.9% at 5 pixels, with a 141-millisecond runtime. On RDNIM, the figures were 1.69, 58.0%, 43.6% and 83 milliseconds. The authors state that UPAL had the highest repeatability among the evaluated methods.
What the tests add up to
UPAL was trained with a teacher heatmap formed from SuperPoint and DaD predictions, using 10,000 Oxford-Paris distractor images at 800 by 800 pixels. For line processing, its accelerated LSD variant uses a seed-grid stride of 2 and retains the top 20% of pixels; the authors report no performance impact and a threefold efficiency improvement.
A nonrandomized comparison of model variants also favored joint training on MegaDepth. The joint-training version reported AUC values of 58.2%, 70.8% and 79.4% at 5, 10 and 20 degrees, compared with 55.2%, 67.6% and 76.4% for the points-only version. The listed differences were 3.0, 3.2 and 3.0 percentage points, so the comparison does not by itself establish that joint training caused the gap.
In a downstream visual-localization test, the point-and-line configuration on the Stairs scene from 7Scenes reported median translation and rotation errors of 4.4 centimetres and 1.23 degrees. It reached 54.6% accuracy at a threshold of 5 centimetres and 5 degrees. On ETH3D, the reported reconstruction completeness values were 0.30, 1.09 and 4.45 at 1, 2 and 5 centimetres, while accuracy values were 64.80, 77.57 and 88.87.
The evidence has clear boundaries
The results are descriptive: the study reports no inferential statistical tests, confidence intervals or run-to-run variability. Timing and memory comparisons depend on the hardware, implementations and settings used, and baseline defaults may differ.
The main visual-localization result focuses on the Stairs scene because it has the most lines, while other scenes are reported only in supplementary material. Line matching uses descriptors at the endpoints of LSD segments and is reported as vulnerable to viewpoint changes; the study does not test an end-to-end learned point-line matcher.
The document is arXiv preprint version 1, dated 20 August 2026, and says the UPAL code is publicly available on GitHub. Whether the reported balance of accuracy and efficiency holds across more datasets, hardware platforms and independently trained models remains open.
Paper data and sources
Original title: Unified and Efficient Point-Line Local Features
Authors: François Costa, Raphael Kreft, Eckhard Goedeke et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text