SightWarp: Dual Approaches in DTW & XR

• SightWarp is a dual-use research term, referring to both a method that simplifies DTW warping paths by visualizing subsequence distortions and an XR technique that brings distant objects into near-space for direct manipulation.
• In time-series analysis, the method quantifies time-shifts, compression/expansion, and amplitude differences to provide a clearer overview of elastic alignments.
• In head-worn XR, the system leverages natural eye–hand coordination to summon scaled proxies, reducing cognitive load by reconciling far-space object manipulation with direct interaction.

SightWarp is a research name used for two distinct 2025 systems in different technical domains. In time-series analysis, SightWarp denotes a visualization and simplification method for dynamic time warping paths, presented in the paper "Warping and Matching Subsequences Between Time Series," where it is also called Dynamic Subsequence Warping (DSW); its purpose is to identify, quantify, and visualize subsequence-level correspondences and distortions between aligned series (Lin et al., 18 Jun 2025). In head-worn extended reality (XR), SightWarp denotes an interaction technique for direct manipulation of distant virtual objects by exploiting eye–hand coordination to summon scaled near-space proxies of far-space objects to the user’s fingertips (Liu et al., 6 Aug 2025).

1. Terminological scope and domain separation

The term SightWarp therefore has a dual usage rather than a single established meaning. One usage belongs to elastic time-series comparison, where the underlying object is a DTW warping path between two sequences. The other belongs to 3D user interfaces, where the underlying object is a distant virtual object or scene region in head-worn XR.

Usage	Domain	Primary function
SightWarp / DSW	Time-series comparison	Simplifies and visualizes a DTW warping path as subsequence matches
SightWarp	Head-worn XR interaction	Summons a scaled proxy of a distant object for direct hand manipulation

This distinction matters because both systems use the language of “warping,” but the term refers to different technical operations. In the time-series setting, warping is temporal alignment under elastic distance. In XR, warping is the relocation and scaling of a virtual proxy into near space. A plausible implication is that the shared name reflects a common emphasis on preserving structure while making an otherwise hard-to-inspect relation more immediately manipulable.

2. SightWarp as Dynamic Subsequence Warping for time series

In the time-series literature, the SightWarp/DSW method is designed for analysts who have already computed an elastic alignment such as DTW and want to go beyond point-to-point links. The technique supports identifying which contiguous subsequences in series $s_1$ correspond to which contiguous subsequences in series $s_2$, quantifying three distortions between matched subsequences, and visually presenting those subsequences and distortions in a single concise plot (Lin et al., 18 Jun 2025).

The three distortions are time-shift, compression or expansion, and amplitude difference. Time-shift captures one subsequence occurring earlier or later. Compression or expansion captures one subsequence being run faster or slower. Amplitude difference is defined as a constant vertical offset after warping. The central motivation is that traditional visualizations of DTW focus on point-to-point alignment and do not convey broader structural relationships at the level of subsequences. This makes it difficult to understand how and where one time series shifts, speeds up, or slows down with respect to another.

The method takes as input two series $s_1$ and $s_2$ of lengths $n$ and $m$, an optimal DTW warping path $P = [(i_1,j_1),(i_2,j_2),\ldots,(i_L,j_L)]$ with cost matrix $C$, and two tolerances $\delta_{\mathrm{abs}}$ and $\delta_{\mathrm{rel}}$. Its output is a reduced set of breakpoints $s_2$0 defining $s_2$1 piecewise-linear segments that approximate $s_2$2. This design places SightWarp downstream of an already computed elastic alignment, rather than replacing DTW itself.

3. Algorithmic structure, quantitative measures, and visualization in the time-series method

The algorithm proceeds in two phases. Phase 1 is a recursive cost-based RDP procedure on the warping path. It initializes a queue $s_2$3, maintains an empty set $s_2$4 of accepted segment endpoints, and for each interval $s_2$5 compares the optimal-path cost on $s_2$6 with the cost of a straight-line approximation rasterized by Bresenham’s algorithm. The linear approximation is accepted if

$s_2$7

where $s_2$8 and $s_2$9 is the total path length. Otherwise, the point on $s_1$0 farthest in Euclidean distance from the line is used to split the interval recursively. At termination, sorting $s_1$1 and mapping back to $s_1$2 yields a piecewise-linear path $s_1$3. Phase 2 performs local merging by testing whether adjacent segments can be merged while still satisfying the same local tolerance criterion (Lin et al., 18 Jun 2025).

For one simplified segment mapping $s_1$4 to $s_1$5, the paper defines the absolute time-shift $s_1$6 as

$s_1$7

A positive $s_1$8 means $s_1$9 is delayed by $s_2$0 steps relative to $s_2$1; a negative value means $s_2$2 leads. Absolute and relative compression are

$s_2$3

Here, $s_2$4 or $s_2$5 indicates compression, while $s_2$6 or $s_2$7 indicates expansion. The amplitude-difference envelope is defined on the rasterized matched pairs $s_2$8 by

$s_2$9

$n$0

so that the warped amplitude of point $n$1 in $n$2 lies in $n$3.

The method also provides a global distance error bound. If $n$4 is the original DTW distance and $n$5 is the distance along the simplified path, then

$n$6

Specializing $n$7 or $n$8 recovers purely absolute or purely relative guarantees.

The visualization overlays the two time-series plots on a shared time axis. Each breakpoint pair is connected by a straight chord, and the $n$9 chords replace the full cloud of point-to-point lines. Chords are drawn at endpoints only, so subsequence matches appear as contiguous bands. An orange horizontal bar of length $m$0 encodes absolute compression over each chord’s horizontal span, and an orange shading band from $m$1 to $m$2 shows amplitude difference on the $m$3 curve. A small arrowhead or text label near each chord can show $m$4 numerically on hover. Interactive sliders allow variation of $m$5 and $m$6 so that segmentation can coarsen or refine in real time.

The reported examples come from the UCR archive. On ECGFiveDays, DSW collapses a noisy extreme warp at the end into a single segment and shows two large peaks as matched subsequences, with one peak wider and higher. On the UMD dataset, the paper contrasts the use of only relative tolerance with combined tolerance, arguing that the latter yields coarser and more meaningful segmentation. On CBF and TwoPatterns, a large overall shift plus compression across the bulk of the series is captured in one segment, whereas point-to-point visualizations create cluttered crossing lines. In a paused-process example, when $m$7 is held constant for many steps in a $m$8 warp, DSW yields a single horizontal-warp segment rather than dozens of individual links.

4. SightWarp as an XR interaction technique for distant-object manipulation

In XR, SightWarp is an interaction technique for head-worn systems that lets users warp a distant virtual object into immediate reach by exploiting natural eye–hand coordination patterns. It is motivated by a limitation of direct grabbing, which works within arm’s reach, and by shortcomings of indirect methods such as gaze-and-pinch, which do not provide the same proprioceptive feedback as direct gestures. SightWarp is intended to make direct gestures available for far-away objects without extra mode switches or new gestures to learn (Liu et al., 6 Aug 2025).

The technique consists of two complementary pathways. In GazeToHand, after the user has fixated on and pinched a distant object, keeping the pinch engaged, the user looks down toward the pinching hand. This gaze shift signals the system to summon not only the selected object but also its local context to the fingertips. The summoned proxy appears from a new viewpoint, such as a top-down view, anchored at the pinch location. In HandToGaze, the user maintains gaze on a distant object without pinching and brings the free hand into view along the gaze ray. This hand intrusion acts as a trigger for warp. The system then warps the gazed-at region, preserves the original viewing angle, and brings it in front of the hand for direct manipulation.

In both pathways, the proxy remains active as a semi-transparent sphere containing the object and its immediate neighbors until the eye–hand alignment is broken, after which the scene is restored to its original configuration with any transformations applied. A common misconception would be that SightWarp is simply another indirect selection method. The paper instead positions it as a way to recover direct manipulation for far-space objects by inserting a near-space proxy into the interaction loop.

5. System pipeline, geometric formulation, and trigger logic in XR

The XR implementation uses eye tracking at 30 Hz on the Meta Quest Pro. A fixation is detected when the 2D gaze point on the virtual image plane remains within a $m$9 visual angle radius for at least 100 ms. Proxy creation then depends on distinct trigger conditions for the two pathways (Liu et al., 6 Aug 2025).

For GazeToHand, the precondition is that the user has initiated Gaze+Pinch on a far-space object and is holding the pinch gesture. The system continuously computes the 3D angle $P = [(i_1,j_1),(i_2,j_2),\ldots,(i_L,j_L)]$0 between the gaze vector $P = [(i_1,j_1),(i_2,j_2),\ldots,(i_L,j_L)]$1 and the vector $P = [(i_1,j_1),(i_2,j_2),\ldots,(i_L,j_L)]$2 from the eye to the hand:

$P = [(i_1,j_1),(i_2,j_2),\ldots,(i_L,j_L)]$3

When $P = [(i_1,j_1),(i_2,j_2),\ldots,(i_L,j_L)]$4, with $P = [(i_1,j_1),(i_2,j_2),\ldots,(i_L,j_L)]$5, proxy summoning is triggered. A slightly relaxed threshold $P = [(i_1,j_1),(i_2,j_2),\ldots,(i_L,j_L)]$6 is used to prevent accidental releases.

For HandToGaze, the precondition is that the user is gazing at an object without pinching. As the hand moves within the field of view, its 3D position must satisfy two simultaneous constraints: a distance constraint, $P = [(i_1,j_1),(i_2,j_2),\ldots,(i_L,j_L)]$7, and the same angular alignment condition $P = [(i_1,j_1),(i_2,j_2),\ldots,(i_L,j_L)]$8 with $P = [(i_1,j_1),(i_2,j_2),\ldots,(i_L,j_L)]$9. When both are met, the scene proxy is summoned.

The proxy transform is computed from the original object transform $C$0 and the hand pose $C$1. The uniform scale factor is

$C$2

where

$C$3

With $C$4, one expression for the proxy transform is

$C$5

The stated intuition is that this brings the far-space object into the hand’s coordinate frame and uniformly scales it so that its visual size matches the original. For GazeToHand, the same formula applies, although an application can choose to tweak $C$6, for example to enlarge the context sphere for greater precision.

The paper also states threshold equations in more schematic form. For GazeToHand, a gaze-shift threshold

$C$7

ensures that the user intentionally looked down. For HandToGaze, the hand-entry threshold is

$C$8

6. Evaluation, application scenarios, and limitations across the two usages

The XR paper evaluates SightWarp in a 12-participant within-subject study using a standardized 6 DOF docking task that compares Gaze+Pinch, GazeToHand, and HandToGaze. Independent variables are Rotation Magnitude, with $C$9 versus $\delta_{\mathrm{abs}}$0, and Object Size, with $\delta_{\mathrm{abs}}$1 versus $\delta_{\mathrm{abs}}$2 visual angle. Each participant performed 144 trials, corresponding to 3 techniques, 2 rotations, 2 sizes, and 12 repetitions. Metrics included Trial Completion Time, Acquisition Time, First Manipulation Duration, Clutch Count, Failed Gesture Count, Hand Translation and Rotation, NASA-TLX workload, and subjective preference (Liu et al., 6 Aug 2025).

The key reported results are that both GazeToHand and HandToGaze were significantly faster than Gaze+Pinch for Trial Completion Time, with $\delta_{\mathrm{abs}}$3. Gaze+Pinch incurred more re-grabs than the two warp methods for Clutch Count, with $\delta_{\mathrm{abs}}$4. Gaze+Pinch also saw more failed pinches than GazeToHand at $\delta_{\mathrm{abs}}$5 and HandToGaze at $\delta_{\mathrm{abs}}$6. Total hand translation and rotation were reduced in GazeToHand versus Gaze+Pinch, with translation at $\delta_{\mathrm{abs}}$7 and rotation at $\delta_{\mathrm{abs}}$8. First Manipulation Duration was shortest in GazeToHand, with $\delta_{\mathrm{abs}}$9. No significant differences were found in acquisition time, which the authors interpret as indicating that users could summon proxies with minimal overhead. Subjectively, participants rated both warp methods as less fatiguing and easier than pure gaze-and-pinch, though some noted additional eye strain from repeatedly refocusing between distances.

The application scenarios enumerated for XR are 6 DOF Manipulation, Overview-and-Detail Navigation, World-in-Miniature, Zoom Lens interfaces, and Details-On-Demand. In these examples, the proxy can support top-down editing of a neighborhood block, interactive manipulation of a context sphere around a selected region, magnification of distant menus or HUD elements, and appearance of fine labels or annotations only when objects are summoned into near-space proxy form.

The time-series paper presents a different kind of evaluation. Its quantitative claim is the distance guarantee

$\delta_{\mathrm{rel}}$0

and its qualitative comparison emphasizes robustness to noise overfitting. The paper states that, unlike standard DTW visualizations or RDP-based simplification, which is purely spatial, DSW will collapse noisy jagged warps if they add little cost. It also contrasts the method with path-constrained DTW variants, stating that such variants enforce diagonality everywhere and thus cannot show genuine large compressions or pauses. By contrast, SightWarp/DSW works downstream of any warping path and remains compatible with AMDTW, sakoe-chiba windows, and related constraints. Case-study users reportedly found that the $\delta_{\mathrm{rel}}$1-segment view made high-level distortions easier to detect at a glance than a traditional point-link view.

The principal limitations are domain-specific. For XR, the paper identifies parameter tuning for angular alignment, depth range, proxy-sphere size, and CD gain; the need for further studies on complex multi-step workflows, selection tasks, and bimanual AR scenarios; possible unintended triggers or premature release; and spatial conflicts when the proxy overlaps existing geometry. For the time-series method, the paper’s emphasis on interactive tolerance tuning and case-study interpretation suggests that utility depends on the analyst’s choice of $\delta_{\mathrm{rel}}$2 and $\delta_{\mathrm{rel}}$3, although the method preserves an explicit error bound.

Taken together, the two systems show that “SightWarp” is not a single research lineage but a shared label for two efforts to reduce the cognitive burden of interpreting difficult correspondences. In one case, the correspondence is a dense DTW path compressed into semantically meaningful subsequence matches. In the other, it is the eye–hand relation used to convert far-space manipulation into near-space direct action. A plausible implication is that the shared name reflects a common design objective: making latent structure visible and operational at the scale of human action.