Localization Capture Time in Theory & Applications

Updated 8 July 2026

Localization capture time is a multifaceted concept quantifying the rounds needed to deduce an unknown’s position in graph games or jointly infer capture time and location in imaging.
In graph theory, it measures the minimum rounds for cops to determine an invisible robber’s exact position via distance probes, with established linear bounds and extremal cases for various graph classes.
In computer vision and related fields, it refers to the joint inference of temporal and geospatial metadata, enhancing timestamp verification, event localization, and detection in diverse technical domains.

Localization capture time is an overloaded technical notion whose meaning depends on disciplinary context. In graph theory, it denotes the round complexity of the localization game, where an invisible robber is identified through distance probes rather than by co-occupation of a vertex. In geotemporal computer vision, it denotes estimation or verification of an image’s time of capture jointly with geographic localization. Related formulations also appear in temporal action localization, relativistic particle detection, and nonequilibrium biophysical transport, where localization is expressed through the timing of a detection or capture event rather than through a purely spatial observable.

1. Graph-theoretic localization capture time

In the localization game on a connected graph $G$ , the robber is invisible, chooses an initial vertex first, and in each round the cops choose $k$ vertices to probe, receive the distance vector

$(d(c_1,r),\ldots,d(c_k,r)),$

and then the robber moves to a neighbor or stays put. The cops win when the distance information determines the robber’s exact position after finitely many rounds. The minimum number of cops that guarantees eventual localization is the localization number $\zeta(G)$ . For $k\ge \zeta(G)$ , the $k$ -localization capture time $\lcapt_k(G)$ is the minimum number of rounds needed for $k$ cops to win under optimal play; when $k=\zeta(G)$ , this is written $\lcapt(G)$. The number of rounds is explicitly the number of probes (Behague et al., 2021, Chenoweth et al., 14 Aug 2025).

A central conjecture is the Localization Capture Time Conjecture, which asks whether all connected graphs are well-localizable, meaning that there exists a constant $k$ 0 such that every graph $k$ 1 in the family satisfies

$k$ 2

A stronger form asks whether every connected graph on $k$ 3 vertices satisfies $k$ 4. The 2021 formulation introduced localization capture time as a new graph parameter and proved linear bounds for trees, interval graphs, complete multipartite graphs, and incidence graphs of projective planes, while also developing bounds via pathwidth and treewidth (Behague et al., 2021).

The same line of work distinguishes resource complexity from round complexity. Increasing the number of cops beyond $k$ 5 yields temporal speed-up: $k$ 6 decreases with $k$ 7, and if $k$ 8 equals the metric dimension then $k$ 9. This makes localization capture time a minimax timing parameter rather than merely a variant of metric dimension (Behague et al., 2021).

2. Exact bounds, extremal cases, and structural methods

Subsequent work sharpened the tree case substantially. If $(d(c_1,r),\ldots,d(c_k,r)),$ 0 is a tree that does not contain a copy of $(d(c_1,r),\ldots,d(c_k,r)),$ 1, then

$(d(c_1,r),\ldots,d(c_k,r)),$ 2

where $(d(c_1,r),\ldots,d(c_k,r)),$ 3 is the number of leaves. If $(d(c_1,r),\ldots,d(c_k,r)),$ 4 contains a copy of $(d(c_1,r),\ldots,d(c_k,r)),$ 5, then

$(d(c_1,r),\ldots,d(c_k,r)),$ 6

For a tree on $(d(c_1,r),\ldots,d(c_k,r)),$ 7 vertices, this yields

$(d(c_1,r),\ldots,d(c_k,r)),$ 8

The bounds are tight: for the star $(d(c_1,r),\ldots,d(c_k,r)),$ 9, $\zeta(G)$ 0 and $\zeta(G)$ 1, while a family $\zeta(G)$ 2 obtained from $\zeta(G)$ 3 attains the two-cop bound (Chenoweth et al., 14 Aug 2025).

For outerplanar graphs with localization number $\zeta(G)$ 4, two-cop capture time is also linear. If $\zeta(G)$ 5 is an outerplanar graph that is an edge-disjoint union of blocks $\zeta(G)$ 6, then

$\zeta(G)$ 7

If $\zeta(G)$ 8 is 2-connected outerplanar with $\zeta(G)$ 9 chords, then

$k\ge \zeta(G)$ 0

These results establish that outerplanar graphs with $k\ge \zeta(G)$ 1 are well-localizable and satisfy the stronger $k\ge \zeta(G)$ 2-vertex bound (Chenoweth et al., 14 Aug 2025).

Earlier results remain important for broader graph classes. Trees are well-localizable; graphs with $k\ge \zeta(G)$ 3, including interval graphs, satisfy $k\ge \zeta(G)$ 4; and for the incidence graph of a projective plane of order $k\ge \zeta(G)$ 5,

$k\ge \zeta(G)$ 6

The Heawood graph has exact localization capture time $k\ge \zeta(G)$ 7 (Behague et al., 2021).

Graph class	Capture-time statement	Source
Trees without $k\ge \zeta(G)$ 8	$k\ge \zeta(G)$ 9	(Chenoweth et al., 14 Aug 2025)
Trees with $k$ 0	$k$ 1	(Chenoweth et al., 14 Aug 2025)
Outerplanar, $k$ 2	$k$ 3	(Chenoweth et al., 14 Aug 2025)
2-connected outerplanar	$k$ 4	(Chenoweth et al., 14 Aug 2025)
Interval graphs	$k$ 5	(Behague et al., 2021)
Projective-plane incidence graphs	$k$ 6	(Behague et al., 2021)

A further structural development is a coloring-based game structure. For distance colorings, the number of rows in this structure equals $k$ 7, reframing localization capture time as the height of a layered combinatorial object rather than only as a game-theoretic runtime (Chenoweth et al., 14 Aug 2025).

3. Distinction from ordinary capture time in visible pursuit

Localization capture time should be distinguished from the ordinary capture time of the standard visible Cops and Robber game. In that game, the cops and robber occupy vertices of a finite, simple, connected, undirected graph; the cops choose starting vertices first, the robber chooses her start, and thereafter each player may stay put or move to an adjacent vertex. Capture occurs when a cop occupies the robber’s vertex after a cop move. If $k$ 8 is $k$ 9-cop-win, the $\lcapt_k(G)$0-capture time $\lcapt_k(G)$1 is the minimum number of rounds needed for $\lcapt_k(G)$2 cops to force capture against an evasive robber (Mehrabian, 2010).

For Cartesian products of two trees $\lcapt_k(G)$3, the exact visible-game result is

$\lcapt_k(G)$4

and for the $\lcapt_k(G)$5 grid,

$\lcapt_k(G)$6

The paper establishing these formulas is explicit that it is not about localization games, localization number, metric dimension, or locating an invisible robber through probes. Its subject is the standard visible pursuit game, not localization capture time in the graph-theoretic sense (Mehrabian, 2010).

This distinction matters because the two notions use different observables. Ordinary capture time measures how long visible pursuit takes under edge-constrained motion. Localization capture time measures how long it takes to infer the position of an invisible robber from distance vectors, with the cops free to jump between probe vertices each round. The shared phrase “capture time” therefore conceals a substantial change in game model.

4. Geotemporal image inference: predicting when and where an image was captured

In computer vision, localization capture time refers to inference of an image’s time of capture jointly with geographic location. An early metadata-supervised approach learned “geo-temporal image features” by optimizing four tasks: $\lcapt_k(G)$7 Time was parameterized as one-hot month and one-hot hour of day, for a total of $\lcapt_k(G)$8 dimensions, with all timestamps in Greenwich Mean Time; location was encoded as normalized 3D ECEF coordinates; location prediction used $\lcapt_k(G)$9 equal-angle latitude-by-longitude bins, and time prediction used $k$ 0 month-by-hour bins. The resulting representation correlated more strongly with transient outdoor attributes than ImageNet-pretrained features, with average maximum absolute correlation $k$ 1 versus $k$ 2 for ImageNet features and $k$ 3 for random features (Zhai et al., 2019).

A more recent formulation, GT-Loc, makes the coupling between location and capture time explicit. From a single outdoor image, it predicts GPS coordinates together with time-of-day and time-of-year, but not the absolute year, which is explicitly discarded. GT-Loc is retrieval-based rather than a direct timestamp regressor or a coarse classifier: an image embedding is matched against galleries of time embeddings and location embeddings in a shared 512-D space using cosine similarity. The model uses a frozen CLIP ViT-L/14 image backbone with a trainable MLP projection, a GeoCLIP-style location encoder based on Equal Earth Projection and multi-scale Random Fourier Features with $k$ 4, and a parallel time encoder that maps Unix timestamps to normalized cyclic coordinates $k$ 5, where $k$ 6 traces the year cyclically and $k$ 7 traces the day cyclically (Shatwell et al., 14 Jul 2025).

The key temporal modeling choice is toroidal cyclic time. GT-Loc argues that standard hard contrastive learning is structurally inappropriate for time because adjacent hours and neighboring months often look extremely similar in outdoor imagery, while time is periodic. Its Temporal Metric Learning therefore replaces one-hot positives and negatives with soft targets inversely proportional to temporal difference on a toroidal manifold. On zero-shot SkyFinder time prediction, the time-only variant TimeLoc obtained month error $k$ 8, hour error $k$ 9, and TPS $k=\zeta(G)$ 0, whereas joint GT-Loc obtained month error $k=\zeta(G)$ 1, hour error $k=\zeta(G)$ 2, and TPS $k=\zeta(G)$ 3. The time-loss ablation further showed CLIP loss at $k=\zeta(G)$ 4, TML with Euclidean $k=\zeta(G)$ 5 at $k=\zeta(G)$ 6, and TML with cyclic/toroidal distance at $k=\zeta(G)$ 7. Geo-localization remained competitive, with 1 km recall $k=\zeta(G)$ 8 on Im2GPS3k and $k=\zeta(G)$ 9 on GWS15k (Shatwell et al., 14 Jul 2025).

Taken together, these results establish a concrete technical meaning of localization capture time in outdoor vision: the time-of-capture problem becomes more accurate when learned jointly with localization, especially when time is modeled as a cyclic variable rather than as a linear regressor or as a hard class label (Zhai et al., 2019, Shatwell et al., 14 Jul 2025).

5. Timestamp verification, benchmark design, and geo-temporal reasoning

A complementary line of work treats capture time as a consistency variable rather than a direct prediction target. In content-aware timestamp manipulation detection, the input is a tuple $\lcapt(G)$0 consisting of a ground-level image $\lcapt(G)$1, an alleged timestamp $\lcapt(G)$2, a geographic location $\lcapt(G)$3, and an optional co-located satellite image $\lcapt(G)$4. The model estimates

$\lcapt(G)$5

where $\lcapt(G)$6 denotes a consistent tuple and $\lcapt(G)$7 an inconsistent one. Time is represented only by month and hour (UTC), each scaled to $\lcapt(G)$8; location is encoded continuously using ECEF coordinates divided by the Earth’s radius; and the satellite image provides geographic context rather than temporal evidence. On the CVT benchmark, the best model reached $\lcapt(G)$9 accuracy and $k$ 00 AUC, compared with $k$ 01 accuracy and $k$ 02 AUC for Salem et al. The DenseNet ablation showed $k$ 03 at $k$ 04 accuracy and $k$ 05 AUC, $k$ 06 at $k$ 07 and $k$ 08, $k$ 09 at $k$ 10 and $k$ 11, and $k$ 12 with transient attributes at $k$ 13 and $k$ 14. The same model can be repurposed for missing-timestamp estimation by evaluating all candidate month/hour pairs and selecting the maximizer of $k$ 15 (Padilha et al., 2021).

Benchmarking work on MLLMs shifts the emphasis from retrieval accuracy to interpretable geo-temporal reasoning. GTPred introduces a benchmark of 370 globally distributed images spanning over 120 years, with location represented as a hierarchy

$k$ 16

and year labels given as either exact years or intervals. The evaluation uses an interval-aware year score and a hierarchical weighted location score with

$k$ 17

Reasoning quality is separately judged by GPT-5.1 on a 1-to-10 scale and normalized to $k$ 18. The strongest model was Gemini 3 Pro Preview, with time answer $k$ 19, time reasoning $k$ 20, location answer $k$ 21, and location reasoning $k$ 22. The main ablation is asymmetric: removing temporal information hurts geo-localization substantially, while removing geographic information has little effect and can even slightly help time prediction for several models (Li et al., 19 Jan 2026).

These two strands define a broader evaluation landscape for localization capture time. One asks whether a claimed time-of-capture is plausible at a given location; the other asks whether a model can jointly infer year and place with interpretable reasoning. Both treat temporal inference as an integral component of localization rather than as auxiliary metadata (Padilha et al., 2021, Li et al., 19 Jan 2026).

6. Other technical meanings: narrated actions, relativistic detection, and filament-tip capture

In video understanding, the phrase appears as temporal localization of when an event actually happened. The task studied in narrated-action localization is to decide whether a transcript-extracted action is visible and, if so, identify its temporal start and end in the clip. The proposed 2SEAL method first predicts whether an action is short ( $k$ 23 s) or long ( $k$ 24 s); short actions are localized by transcript alignment, while long actions are localized by a multimodal visual-language scorer over overlapping three-second spans with stride $k$ 25 s. On the test set, the best system, 2SEAL + MPU, achieved visibility accuracy $k$ 26 and mIoU $k$ 27, compared with $k$ 28 and $k$ 29 for MPU alone. This usage concerns event time within a video rather than geolocation, but it preserves the same core question: when did the observed event occur (Ignat et al., 2022).

In relativistic quantum field theory, localization is reformulated in terms of time-of-arrival observables. Instead of asking for a particle’s position at a fixed time, the framework asks when a detector at a fixed place is triggered and with what spatially localized record. The resulting probabilities are linear functionals of QFT correlation functions,

$k$ 30

with $k$ 31 the spacetime localization of the detection record. For scalar fields, the paper identifies several time-of-arrival observables that differ by how the apparatus localizes detection records, and shows that maximum localization is obtained for a unique observable related to the Newton–Wigner position operator. It also derives a detector-independent lower bound on the variance of capture time,

$k$ 32

Here “capture time” is literally the time coordinate of a detector event (Anastopoulos et al., 2018).

In biophysical transport on dynamic filaments, localization by capture has yet another meaning. A lattice-gas model for microtubule-binding proteins such as XMAP215 and MCAK studies diffusion on a semi-infinite one-dimensional lattice whose terminal site is a tip reaction site. Capture is implemented by forbidding hopping from site $k$ 33 back to site $k$ 34, thereby breaking detailed balance. The central localization observables are tip occupancy $k$ 35 and the capture flux

$k$ 36

with steady-state tip balance

$k$ 37

The paper does not compute an explicit mean first-passage capture time, but tip occupancy, capture flux, and tip residence time $k$ 38 act as kinetic proxies. The main result is that nonequilibrium diffusion plus capture strongly enhances tip localization relative to direct binding from solution alone (Reithmann et al., 2016).

Across these usages, the common pattern is not a single universal definition but a shared shift from purely spatial localization to localization mediated by time: rounds of probing in graphs, time-of-capture estimation in images, temporal intervals in video, detector arrival times in QFT, and capture-mediated residence at a reaction site in biophysics.