LR-Distance: A Multi-Domain Distance Framework

Updated 10 July 2026

LR-Distance refers to a collection of context-specific distance constructs used in vision, geometry, and learning, each defined with distinct methodologies and metrics.
It underpins key applications such as long-range distance prediction in autonomous driving, panoramic localization via spherical line functions, and depth recovery in 3D detection.
Its diverse implementations deliver robust performance, exemplified by metrics like 38.5% error threshold achievement in object distance estimation and enhanced mAP in long-range detection.

LR-Distance is not a single standardized construct across the research literature. In the supplied arXiv usage, it denotes several distinct but technically related ideas: Long-Range Distance Estimation for small or distant objects in vision systems, line distance functions on the sphere for panoramic localization, Lorentzian distance-squared mappings in Minkowski space, and distance-ratio-based posterior models in metric learning. The shared theme is the use of distance as a structured computational primitive rather than as a generic scalar similarity score (Li et al., 2022, Kim et al., 2023, Ichiki et al., 2013, Kim et al., 2022).

1. Terminological scope and principal senses

The term is used heterogeneously across computer vision, localization, geometry, and learning theory. In autonomous driving and aerial perception, “LR” is tied to long-range sensing and range estimation. In panoramic localization, it is tied to line distance functions. In differential geometry, it refers to Lorentzian distance-squared mappings. In metric learning, it appears as a distance-ratio formulation whose log-odds depend on ratios of distances (Li et al., 2022, Kim et al., 2023, Ichiki et al., 2013, Kim et al., 2022).

Area	Meaning of LR-Distance	Representative source
Autonomous driving and aerial vision	Long-Range Distance Estimation or long-range range prediction	(Li et al., 2022, Zhu et al., 2019, Rouhi et al., 5 Aug 2025)
Panoramic localization	Line distance functions on the unit sphere	(Kim et al., 2023)
Lorentzian geometry	Lorentzian distance-squared mappings in Minkowski space	(Ichiki et al., 2013)
Metric learning	Distance-ratio posterior formulation	(Kim et al., 2022)

This distribution of meanings suggests that LR-Distance is best treated as a family of domain-specific constructs rather than a universal term of art.

2. Long-range distance estimation in visual perception

In autonomous driving, Long-Range Distance Estimation is defined as estimating the distances to long-range objects, specifically those beyond typical LiDAR sensing range, with distances measured from the camera to the object center along the camera’s optical axis (Li et al., 2022). R4D addresses this task by using references with known distances in the scene. It builds a graph that connects a target object to all references, defines an edge embedding from target, reference, union, and geo-distance embeddings, and fuses them with an attention module:

$H = \sum_{i=1}^{N} \alpha_i\, h_i.$

Training combines an absolute distance head with auxiliary relative-distance supervision based on

$\Delta d_i = d_t - d_{r,i},$

and Smooth L1 losses for absolute and relative prediction (Li et al., 2022). On the Waymo Open Dataset – Long-Range Labels, R4D reports 38.5\% at $<5\%$ , 62.3\% at $<10\%$ , 77.1\% at $<15\%$ , with Abs Rel 9.8\% and RMSE 26.7 (Li et al., 2022).

A related monocular formulation predicts an object-specific distance directly from an RGB image and a 2D bounding box. That work defines distance as the object’s depth along the camera optical axis, i.e. the camera-frame $Z$ -coordinate, and uses ROI-pooled features with a distance regressor, a classifier, and, in the enhanced model, a keypoint regressor with a projection loss (Zhu et al., 2019). Its projection model is

$s\,u_i = K [R \mid t] X_i,$

and the enhanced loss includes

$L_{3Dpoint} = \frac{1}{N}\sum_i \frac{1}{d_i^*}\left\| P [K(F_i), D(F_i)] - u_i^* \right\|_2.$

On KITTI, the enhanced VGG16 model yields Average RMSE 6.87 m versus IPM 78.87 m and SVR 24.25 m, with the paper emphasizing robustness on curved-road scenes and far-range objects (Zhu et al., 2019).

Long-range distance has also become a dataset-level concern in aerial vision. LRDDv2 introduces 39,516 meticulously annotated images and includes target range information for over 8,000 images, explicitly enabling algorithm development for drone range estimation. The dataset is tailored for long-range aerial object detection, and the majority of its images capture drones with 50 or fewer pixels in 1080p resolution (Rouhi et al., 5 Aug 2025).

3. Long-range 3D detection and perception systems

Long-range distance also appears as a depth-recovery problem inside 3D object detection. LR3D assumes close-range objects have 2D+3D labels while distant objects have only 2D boxes, and learns an implicit inverse projection from 2D box geometry to metric depth (Yang et al., 2024). Its forward and inverse formulations are

$f(d,s,o) = b_{2d}, \qquad f^{-1}(b_{2d}\mid s,o) = d.$

The depth prediction is implemented by a dynamic-weight Implicit Projection Head:

$d_i = f^{(f_g(F_i))}\big(f_{\text{PE}}(b_{2d,i})\big).$

For evaluation, LR3D introduces Long-range Detection Score (LDS) based on relative distance error,

$\Delta d_i = d_t - d_{r,i},$ 0

with mAP over relative thresholds $\Delta d_i = d_t - d_{r,i},$ 1 and

$\Delta d_i = d_t - d_{r,i},$ 2

On KITTI, IP-FCOS3D without distant 3D labels reaches distant LDS/mAP 36.2/31.0\%, approaching fully supervised FCOS3D at 38.2/32.6\% (Yang et al., 2024).

At a larger spatial scale, Self-Supervised Sparse Sensor Fusion for Long Range Perception targets perception distances of at least 250 m for highway driving and heavy trucks, and argues that dense BEV systems scale quadratically with range (Palladin et al., 19 Aug 2025). Its sparse lifting relation is

$\Delta d_i = d_t - d_{r,i},$ 3

and temporal fusion uses motion-aware sparse attention after ego-motion and velocity warping. The reported system extends object detection and forecasting to 250 meters, with 52.61 mAP and 58.06 NDS, a 26.6\% improvement in mAP over the second-best method, and a 30.5\% decrease in Chamfer Distance in LiDAR forecasting (Palladin et al., 19 Aug 2025).

Distance can also act as an explicit architectural partition variable. In LiDAR BEV detection, one study trains separate detectors for close range and long-range objects because point density and local statistics vary strongly with distance (Engels et al., 2020). The KITTI split uses a close-range network for objects with center distance $\Delta d_i = d_t - d_{r,i},$ 4 m and a long-range network for objects with center distance $\Delta d_i = d_t - d_{r,i},$ 5 m in training, with non-overlapping inference regions. The result is improved long-range AP, including +5.0 AP at IoU 0.70 for the Hard category in the 35–70 m range and +14.7 AP at IoU 0.50 in the same regime (Engels et al., 2020).

4. Line distance functions for panoramic localization

In panoramic localization, LR-Distance refers to the line distance functions introduced by LDL for matching panoramas to 3D maps using line geometry (Kim et al., 2023). The panorama is treated on the unit sphere $\Delta d_i = d_t - d_{r,i},$ 6, and 2D line segments are represented by spherical endpoints $\Delta d_i = d_t - d_{r,i},$ 7, $\Delta d_i = d_t - d_{r,i},$ 8. The point-to-line-segment geodesic distance is

$\Delta d_i = d_t - d_{r,i},$ 9

The 2D line distance function is then

$<5\%$ 0

and the 3D counterpart under pose $<5\%$ 1 is

$<5\%$ 2

LDL compares these distributions at uniformly sampled spherical query points using a robust inlier-count loss:

$<5\%$ 3

To resolve rotational ambiguity, the method decomposes line sets along principal directions and sums inlier counts over direction groups. Candidate rotations are generated by triplet matching of 2D and 3D principal directions, the top poses are selected by this line-distance objective, and final refinement uses SuperGlue + PnP-RANSAC (Kim et al., 2023). The method is explicitly designed to avoid one-to-one line correspondences and instead matches the distribution of lines through their distance fields.

5. Lorentzian distance-squared mappings

In differential geometry, LR-Distance denotes Lorentzian distance-squared mappings in Minkowski space $<5\%$ 4, equipped with the Lorentzian inner product

$<5\%$ 5

For a point $<5\%$ 6, the Lorentzian distance-squared function is

$<5\%$ 7

Given $<5\%$ 8, the associated mapping is

$<5\%$ 9

Its $<10\%$ 0-equivalence class is determined by the likeness of the recognition subspace

$<10\%$ 1

For $<10\%$ 2 and points in general position, the classification theorem gives three normal forms (Ichiki et al., 2013). If $<10\%$ 3 is time-like,

$<10\%$ 4

where

$<10\%$ 5

If $<10\%$ 6 is space-like,

$<10\%$ 7

where

$<10\%$ 8

If $<10\%$ 9 is light-like,

$<15\%$ 0

The Jacobian is

$<15\%$ 1

and for $<15\%$ 2 in general position the singular set coincides with the affine subspace spanned by the base points (Ichiki et al., 2013). In codimension one, the normalized fibers are circles in the definite case, equilateral hyperbolas in the indefinite case, and parabolas in the light-like case.

In metric learning, LR-Distance appears as a distance-ratio posterior model. Given a query embedding $<15\%$ 3, class representative $<15\%$ 4, and distance $<15\%$ 5, the class posterior is

$<15\%$ 6

equivalently a softmax over logits $<15\%$ 7 (Kim et al., 2022). The episode loss is the cross-entropy

$<15\%$ 8

A defining property is exact scale invariance: if all distances are multiplied by $<15\%$ 9, the probabilities and loss are unchanged. In the two-class case, the posterior can be written as

$Z$ 0

and the log-odds become

$Z$ 1

This makes the classifier a direct function of a log-distance ratio (Kim et al., 2022). On CUB with a Conv4 backbone, the reported 1-shot and 5-shot accuracies rise from 44.42 ± 0.84 and 64.24 ± 0.72 for the softmax-based ProtoNet to 57.13 ± 0.95 and 76.50 ± 0.66 for the DR formulation (Kim et al., 2022).

A common terminological confusion is that not every paper with “LR” in its title concerns LR-Distance. In the algebraic theory of LR pairs and LR triples, the term “LR-Distance” does not appear; the paper instead studies lowering and raising linear transformations, and identifies the parameter sequence $Z$ 2 as the closest distance-like numerical invariant (Nomura, 2015). This distinction is important because it separates distance-centric usages from purely algebraic LR notation.

Across these literatures, LR-Distance is therefore best understood as a context-dependent formal object: a long-range range predictor in perception, a spherical line-distance field in localization, a Lorentzian quadratic mapping in geometry, or a scale-invariant distance-ratio posterior in metric learning.