TimeLineXR: An Interactive Exploration of Time in Vision-Language Models

I. INTRODUCTION

• VLMs encode time along a structured, low-dimensional axis, but users cannot see or query that structure interactively.The paper frames exploration as a step toward interpretability of VLM representations.Source: §I · p. 2
• TimeLineXR exposes the VLM time axis through a browser demo and a room-scale WebXR scene that users can manipulate and query with images.The WebXR interface supports grab, rotate, and scale interactions, plus throw-to-time queries.Source: §I · p. 2
• The browser demo lets users upload or select images and view their estimated first-appearance years on an interactive 3D time axis.The demo also includes a running prediction table.Source: §I · p. 2
• The WebXR interface renders the same time axis immersively at room scale, with grab, rotate, scale, and throw-to-time interactions.Source: §I · p. 2
• A shared backend packages the CLIP, cosine-kernel KPCA, Bézier-axis, and nearest-point estimation method for both clients.The web demo uses a server-side Gradio Space, while the WebXR client uses a FastAPI HTTP service.Source: §I · p. 2

II. RELATED WORK

• The reference method finds a VLM time axis approximable by a 1D Bézier curve in a kernel PCA subspace, while earlier visual dating focused on historical photographs in one domain,.Source: §II · p. 2
• Unlike a generic embedding viewer, TimeLineXR provides a temporal index and lets users query, place images on, and browse along the time axis.Embedding Projector is described as a generic projection viewer without time estimation capabilities [2].Source: §II · p. 2
• TimeLineXR combines a VLM-derived time axis, live image placement, and WebXR interaction, drawing on immersive analytics and cultural-heritage contexts,.The paper states that no prior demonstration combines these elements in one package.Source: §II · p. 2

III. SYSTEM OVERVIEW

• The fitted Bézier time axis and KPCA projection state are computed once offline and reused for every query.Source: §III · p. 2
• Web uploads and WebXR throw-to-time interactions use the same estimation pipeline and return an estimated year plus a 3D time-axis coordinate.Source: §III · p. 2
• The rendered axis uses per-year text embeddings reduced by KPCA and projected onto the fitted curve, placing each query beside the years that best match it in embedding space.Source: §III · p. 2

A. Time Estimation Backend

• Year prompts are encoded, reduced with cosine-kernel KPCA, and fitted by a smooth 1D Bézier curve, arranging years from 1700 to 2024.Source: A. Time Estimation Backend · p. 2
• At query time, the image is encoded by the same VLM, projected into the same subspace, and assigned the year of its nearest point on the curve.Source: A. Time Estimation Backend · p. 2
• Using a 3D projection instead of the reference 13-dimensional subspace increases MAE, a trade-off the paper accepts for spatial legibility.Users can switch between CLIP ViT-B/32 and EVA-CLIP in the web demo to observe how backbone choice affects the estimated year.Source: A. Time Estimation Backend · p. 2

B. Web Demo

• The Gradio web demo renders the time axis as an interactive Plotly 3D scatter.Source: B. Web Demo · p. 2
• The two-phase pipeline builds and caches a time axis offline, then encodes each query image with CLIP and snaps it to the axis to return an estimated year.Source: Figure 2 · p. 3
• Table I reports a TIME10k MAE dimensionality trade-off, comparing S=13 in the reference with S=3 in TimeLineXR for a B´ezier(RS, NN) variant of.The supplied passages do not state the MAE values.Source: Table I · p. 3, B. Web Demo · p. 3
• Users upload an image or choose a bundled example; its estimated year appears on the plot and in a running prediction table.Year embeddings appear as color-coded points on a continuous red B´ezier curve, and the browser demo is usable without a headset.Source: B. Web Demo · p. 3

C. WebXR Application

• The time axis appears as a room-scale point cloud with a B´ezier tube, and users can grab, move, and pinch-scale the scene.Source: C. WebXR Application · p. 3
• Desktop browsers provide an orbitable preview; headsets request passthrough mixed reality first and fall back to a dark virtual environment.Source: C. WebXR Application · p. 3
• Users throw a card from a grabbable shelf or upload an image; the card settles at the estimated year on the time axis.Source: C. WebXR Application · p. 3
• The throw animation lasts 0.5–1.3 seconds to cover typical backend prediction times on CPU and GPU hosts, masking latency while giving the placement a physical feel.The demonstration video was recorded on a Meta Quest 3; uploaded and bundled images are the universal input path because headset browsers do not reliably expose raw passthrough-camera access.Source: C. WebXR Application · p. 3

IV. DEMONSTRATION SCENARIO

• The time axis exposes structure that correlation scores and error metrics do not show, including its shape, dense and sparse regions, and image attachments.Source: §IV · p. 3
• Users can add mobile phones across decades, compare objects across categories over the same period, or load a TIME10k axis with hundreds of images from all six categories.These interactions let users qualitatively examine ordering and whether images group by era or domain, then add their own images to probe beyond the benchmark.Source: §IV · p. 3
• In the Meta Quest 3 XR version, walking along and rotating the room-scale point cloud makes the axis's non-linearity tangible in a way a 2D plot cannot.Source: §IV · p. 3

V. IMPLEMENTATION NOTES

• The backend uses frozen pretrained encoders and introduces no new training.Its stack includes Python, PyTorch, HuggingFace Transformers, and scikit-learn.Source: §V · p. 3
• The browser demo's left panel specifies the embedding backbone and query path, the center shows the year-colored 3D time axis and query placement, and the right provides bundled examples.Source: Figure 3 · p. 4
• The web client uses Gradio and Plotly, while the WebXR client uses FastAPI/uvicorn and Three.js; immersive WebXR sessions require HTTPS.Source: §V · p. 4
• Table II reports typical backend prediction latency on CPU-only and GPU-enabled hosts alongside the XR throw-animation duration.The reported backend timing covers CLIP encoding, KPCA projection, and B´ezier snapping; the supplied passage gives no timing values.Source: §V · p. 4, Table II · p. 4

Technical

• Current headset access does not reliably expose raw passthrough-camera feeds; the authors plan to add live-camera queries if that changes.The planned workflow would place a physical object's position on the time axis without an upload step.Source: §VI · p. 4
• Lighter on-device backbones such as MobileCLIP could make the planned live-camera query loop fully local.Source: §VI · p. 4

Applicational

• The authors plan museum and exhibition deployments where visitors follow a collection and estimate each item's year while seeing decade-by-decade changes in appearance.Source: §VI · p. 4

• TimeLineXR combines a browser demo and WebXR scene, built on a shared estimation pipeline, for interactively exploring how a VLM organises time.Source: §VII · p. 4
• Users can query the time axis with their own images and compare where different objects are placed.Source: §VII · p. 4
• The visualization makes the time structure inspectable as a qualitative complement to the method's quantitative results.Source: §VII · p. 4