Large images such as retinal scans, crack inspection imagery and gigapixel slides are too big to segment in one pass, so they are cut into tiles, each tile is segmented on its own, and the results are stitched together. This breaks thin, faint structures (capillaries, hairline cracks, fibers) in a specific way. At the tile edge, the model sees a filament that trails into low contrast with no context on the far side. The next tile sees a faint smudge with no hint that a vessel is arriving. The two halves rarely reconnect. We propose to treat tiled inference the way a pen treats a page. Tiles are visited in an order that follows the structures. Each tile hands its neighbor a compact description of every curve leaving through the shared edge: where it exits, which way it is heading, how it has been bending, and how confident the model is. The neighbor's decoder is conditioned on that description, so it knows where to look and in which direction, which is exactly the information missing when contrast is low.
The temporal-stroke literature supplies the mechanism. Handwriting research has shown that a static ink trace still carries recoverable dynamics: direction, curvature and flow can be summarized by path signatures, which are iterated integrals along a curve (Offline Writer Identification based on the Path Signature Feature). Sequential re-scanning of a static image also recovers cues that static CNNs miss (Static-Dynamic Interaction Networks for Offline Signature Verification). We apply that idea to a place where the "time" axis is real and artificial at once, namely the order in which tiles are processed. Low-contrast ambiguity is the regime where the handoff should matter most, and patch seams are where it is applied.
Core Claim
Our hypothesis is that a low-dimensional, directional curve state passed across tile boundaries recovers seam-crossing connectivity of faint curvilinear structures better than passing more pixels. Overlap halos, larger tiles and global-to-local cascades all supply more pixels. The curve state gives the receiving tile a prior on position and heading that pixels alone cannot provide when the structure is near the noise floor.
Concretely, for each predicted centerline fragment that meets a tile edge, the sending tile emits a handoff token
$$h = \big(p,\ \tau,\ S_{\le 2}(\gamma_{[-L,0]}),\ c\big),$$
with the following parts:
- $p$ is the exit point on the edge.
- $\tau$ is the unit tangent at exit.
- $S_{\le 2}(\gamma_{[-L,0]})$ is the path signature truncated at level two of the last $L$ pixels of the fragment's centerline $\gamma$. Level two includes the signed Lévy area, a cheap summary of how the curve has been turning.
- $c$ is the mean predicted probability along that stretch.
The receiving tile's decoder cross-attends to all incoming tokens on its edges. A separate small trajectory network is trained to continue a curve from its token. It renders a soft "expected corridor" into the neighboring tile, and this corridor is fed as an extra input channel.
The new element is that the prior on curve continuation is learned from data with genuine temporal order. That data includes human pen trajectories from QuickDraw!, Vector-MNIST and FS-COCO, and synthetic L-system curves of the kind CaPro uses. The prior is then fine-tuned on target centerlines. Whether the human-stroke pretraining helps at all is an explicit, falsifiable question in the program.
Relation to Prior Work
Two lines of work attack seams and fine boundaries in very large images. CRM refines coarse masks with an implicit coordinate decoder at arbitrary resolution. CascadePSP pairs a downsampled global pass with sliding-window local refinement. Both supply context by showing the model more of the image, either globally downsampled or locally overlapped. Downsampling is precisely what erases thin, faint filaments, a tension the file notes for SparseMat-style low-resolution gating (SparseMat). Neither method passes any directional description of a structure from one region to the next. BPT replaces the rigid grid with content-clustered tokens and supervises boundaries with Canny edges. Canny targets are unreliable exactly in low-contrast regimes. These works motivate the problem and constrain the design: our base segmenter must run at full resolution on tiles and must not depend on a low-resolution prior to decide where to look.
The temporal-stroke work supplies the method. Path signatures (Offline Writer Identification based on the Path Signature Feature) give a compact, reparameterization-robust description of a curve segment. We use them as a local handoff state rather than as a global writer descriptor, and we compute them on predicted soft centerlines rather than on binarized ink, which that paper requires and which fails at low contrast. SketchINR shows that strokes can be modeled as continuous functions of time and stroke index, decoded in parallel. Its reported pathology of high-frequency jitter warns us to keep the continuation network short-horizon and smooth. SDINet motivates the bet that sequential processing of static imagery recovers information. We differ in that our sequence is a sequence of tiles, and what flows along it is explicit geometry.
CaPro is the closest connection between stroke structure and low-contrast curvilinear segmentation. It uses a codebook learned on MNIST digit shapes to filter prompts for the Segment Anything Model (SAM). CaPro uses stroke shapes statically, on single images, with no tiling. We use stroke dynamics to predict continuation, and only across seams. Iterative refinement by recycling outputs (RecycleNet) suggests a second sweep in the reverse tile order, which we include as an option. Class-Agnostic Visio-Temporal Scene Sketch Semantic Segmentation encodes drawing order to group strokes. It reports sharp drops on unseen categories, which is a warning that stroke priors may not transfer across domains. That is why the transfer question below is decisive rather than assumed.
Research Program
The system has three parts.
The first is a standard full-resolution tile segmenter, a U-Net-class model of the LSSF-Net size regime (LSSF-Net). It outputs a vessel or crack probability map and a soft centerline map.
The second is a handoff module that extracts tokens $h$ at every edge crossing. Its decoder cross-attends to incoming tokens and to a rendered corridor channel.
The third is a scheduler. In the simplest version it sweeps the tile grid in raster order. In a second version it prioritizes tiles that have pending high-confidence incoming tokens, so processing follows the structures the way a pen would. An optional reverse sweep lets tiles revise in light of downstream evidence.
Training simulates tiling on full images, with random tile sizes and offsets so that the seams fall everywhere. The loss is pixel-wise binary cross-entropy plus a soft-centerline connectivity term. The continuation network is pretrained in three variants so that their contribution can be separated: on human pen trajectories, on L-system curves only, and from scratch on target centerlines.
We will also measure where the gain comes from as a function of local contrast. We bin seam crossings by the contrast-to-noise ratio of the ground-truth structure in a band around the seam. The prediction is that the handoff gain grows as contrast falls and vanishes for high-contrast structures. If the gain is flat across contrast, the mechanism is doing something other than what we claim.
Baselines
The naive reference is independent tiling with no overlap and simple concatenation. It sets the floor and defines how much connectivity is lost to seams. It is run with the same segmenter weights as every other condition.
The strongest simple competitor is the same segmenter with overlapping tiles: a halo of 25% or 50% and Gaussian feathering of the probabilities. This is what practitioners actually use. It gives the receiving tile real pixels from across the seam, which is more raw information than our token carries. To keep the comparison fair, the overlap and our method are matched on total pixels processed and wall-clock time, and the overlap width is tuned on validation with the same budget we spend tuning token length $L$.
The strongest published methods for seams and fine boundaries are CRM and CascadePSP, both available through the EntitySeg toolbox for CRM. They are applied as refiners on top of the independent-tile output and retrained on the same curvilinear training splits. The only reported figures are on BIG, which contains objects rather than filaments. There, CRM reports a 15.12% mean Boundary Accuracy gain and a 4.01% IoU gain over coarse masks, and CascadePSP reports 10–15% mBA gains at about 6.6 s per 4K image. No reported result found on Lacuna for CRM or CascadePSP on tiled DRIVE, OCT1600, OCTA-500 or DeepCrack, so we will produce these numbers ourselves.
For reference on the curvilinear benchmarks, CaPro reports 64.33% mIoU on OCT1600 and 58.68% on OCTA-500 (CaPro). That is a single-unlabeled-image setting, so it is context rather than a direct competitor to supervised tiled segmentation.
Two ablations isolate the key ingredient. The first replaces our geometric token with a learned feature vector of the same size, sampled at the exit point. This tests whether explicit curve geometry matters or whether any cross-tile message would do. The second strips the token down to position only, then position plus tangent, then the full signature. This tests whether directional and curvature information is what helps.
Beating the field means the following. On seam-crossing connectivity error at the lowest contrast tercile, the method must clearly beat matched-compute overlap blending. Seam-crossing connectivity error is the fraction of ground-truth centerline crossings of a tile edge where the prediction is disconnected within a small tolerance band. In addition, overall F1 must not degrade, and our method must beat CRM and CascadePSP on the same metric at lower latency.
Why not just…?
Why not just use overlapping tiles with blending? This is the real threat, and the matched-compute comparison above is designed to settle it. Overlap supplies pixels, but at low contrast a faint vessel in the halo is as ambiguous to the receiving tile as it was to the sending one. Feathering then averages two uncertain predictions. CRM reports that blended windows produce smoothed, blobby boundaries. We expect overlap to close most high-contrast seam breaks and our method to win only in the low-contrast bins. If overlap matches us there too, the core claim fails and we will say so.
Why not just bridge gaps afterwards with a minimal-path or skeleton-linking post-process? Geodesic gap closing on the stitched probability map is cheap and directly targets broken filaments, so it is a serious alternative. Its weakness is that it can only link fragments both tiles already detected, and it links by image cost, not by heading. At low contrast the far-side fragment is often absent, and morphological or threshold heuristics discard small seam fragments as noise, as the file notes for size-based polarization (Parallel Res2Net-based Network with ReverseAttention for Polyp Segmentation). We include geodesic linking, applied to both the naive and the overlap outputs, as an additional baseline. We also report how many of our recovered crossings had no far-side fragment at all in the baseline. That count is the part post-processing cannot reach by construction.
Why not just use bigger tiles or a global-to-local cascade? Larger tiles reduce the number of seams but do not remove them, and memory is the reason tiles exist. A downsampled global pass erases exactly the thin, faint filaments at issue, the prior-dependency failure noted for SparseMat. CascadePSP and CRM are run as baselines to test this directly. We also sweep tile size to show how the gap between methods changes as seams become rarer.
Why pretrain on human pen strokes at all, rather than on vessel centerlines? Possibly we shouldn't. CAVT's drops on unseen categories (Class-Agnostic Visio-Temporal Scene Sketch Semantic Segmentation) suggest stroke priors may transfer poorly. The three-way pretraining comparison (human strokes, L-system curves, target only) answers this. If target-only matches the others, the contribution is the handoff mechanism, and the stroke data is dropped from the claim.
Proposed Experiments
The main benchmarks are DRIVE, OCT1600, OCTA-500 and DeepCrack, which are the curvilinear low-contrast sets in the file. All are evaluated under a simulated tiling protocol with tile sizes from 48 to 256 pixels and random grid offsets, so that each image yields many seam crossings. A second protocol degrades contrast by controlled additive noise and intensity compression, which tests the contrast-dependence prediction at a finer resolution than natural variation allows.
Metrics are F1 and mIoU as reported by CaPro, seam-crossing connectivity error, and Betti-number error on tiled crops for topology. Latency is measured on the same GPU. As a sanity check that the method does not harm non-curvilinear boundaries, the base pipeline is also run on BIG, where CRM's reported gains give a reference point.
The decisive checks are three. First, in the lowest-contrast tercile, does the full token beat both overlap blending and geodesic linking on seam-crossing connectivity at matched compute? Second, does the geometric token beat a same-size learned feature message? Third, does the gain shrink toward zero as contrast rises? If all three hold, passing a description of the curve beats passing pixels at the seam. If only the second fails, the result is a cross-tile messaging result, not a trajectory result, and we will report it as that.
References
- Offline Writer Identification Based on the Path Signature Feature
- Static-Dynamic Interaction Networks for Offline Signature Verification
- CaPro: Curvilinear-aware Prompt Learning with Single Unlabeled Image for Cost-effective Curvilinear Structure Segmentation
- High Quality Segmentation for Ultra High-resolution Images
- CascadePSP: Toward Class-Agnostic and Very High-Resolution Segmentation via Global and Local Refinement
- Ultrahigh Resolution Image/Video Matting with Spatio-Temporal Sparsity
- Ultra-High Resolution Segmentation via Boundary-Enhanced Patch-Merging Transformer
- SketchINR: A First Look into Sketches as Implicit Neural Representations
- Abstract: RecycleNet - Latent Feature Recycling Leads to Iterative Decision Refinement
- Class-Agnostic Visio-Temporal Scene Sketch Semantic Segmentation
- LSSF-Net: Lightweight segmentation with self-awareness, spatial attention, and focal modulation
- Parallel Res2Net-based Network with ReverseAttention for Polyp Segmentation