acceptodds
Under review as a conference paper at ICLR 2027

Relative Topology as a Supervisory Signal for Long-Range Dependency

Abstract

Topological data analysis provides principled descriptors of global structure beyond local visual appearance. Standard topological invariants, however, characterise an entire space and are therefore insensitive to which elements are selected by a query, limiting their use for query-dependent relational tasks. We address this mismatch using _relative homology_, which characterises a space with respect to a designated subspace. We introduce _relative topological supervision_, a training-only auxiliary objective in which the query determines the reference subspace and relative homology provides a target encoding the resulting query-dependent structural relation. The target is computed directly from each training image and query, requires no additional annotation, and the auxiliary prediction head is discarded at inference. On the long-range same-path task in Pathfinder, relative topological supervision enables a plain feedforward convolutional network trained from scratch, without residual connections, recurrence, or attention, to solve the task, whereas matched controls using query-independent global topology or permuted topological targets fail. The method remains effective with compact models and limited training data, and we further demonstrate its applicability to retinal vessel segmentation. These results show that relative topology provides a principled mechanism for aligning topological supervision with query-dependent visual reasoning, allowing simple neural architectures to learn global relational structure that is otherwise difficult to capture.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
topological data analysis, long range dependency, pathfinder, relative homology

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_BJBROpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a novel auxiliary objective for the Pathfinder long-range same-path task, derived from relative homology. The task boils down to identify if two endpoints q={q1,q2}q = \{q_1, q_2\} of dashed curves belong to the same curve or not. The authors’ proposed target is δβ0=β0(X)−β0(X∪Aq)\delta \beta_0 = \beta_0(X) - \beta_0(X\cup A_q), where AqA_q is the convex hull of qq, i.e. the line segment that connects both endpoints. This target equals the number of components the line intersects minus 1. The authors predict this number during training in a supervised manner and discard the corresponding head during inference. The main result is presented in Table 1, in which it is shown that a CNN, trained with this auxiliary target, reaches 98.15±0.06%98.15 \pm 0.06 \% test accuracy versus 61.30±6.53%61.30 \pm 6.53 \% with the label-only training. Authors also ablate other auxiliary targets: predicting β0(X)\beta_0(X) alone and permuting δβ0\delta \beta_0 across the training dataset. These auxiliary targets don’t improve the performance, therefore showing that the additional head and the loss term is not enough. Only the correctly paired relative target succeeds, and the authors read this as evidence that the query-anchored topology, and not the extra supervision by itself, drives the gain.

Critical strengths

An observed result is crisp and real. Changing the training loss by adding the proposed auxiliary target allows a plain CNN to perform extremely well on the given task. Using the released code, I reproduced the results of Table 1 for seed 23 and obtained 98.44%98.44 \% accuracy with the proposed loss, while obtaining only 50.67%50.67 \% for the label-only task.

The proposed solution is cheap.: The target for the auxiliary task can be computed from the input and is used only during training, therefore adding nothing to the inference cost.

Critical weaknesses

W1. The topological machinery buys nothing. Authors put relative topology in the title, devote the first figure in text, and spend the whole section explaining the topological construction. However, the gains from the topological machinery are marginal: Figure 1 and the half of the section explain the motivating example, but not the implemented method, the method itself boils down to the calculation of the number of connected components intersected with the interval; the long exact sequence calculation in the Appendix A also gives effectively nothing.

W2. Is it the relative homology that helps? It is unclear if it is exactly the relative 00-th homology that helps the method to excel. Seems clear that the subtraction of 11 can be dropped (because it can be absorbed by the output bias), and the task can be reduced only to the prediction of the number of intersections. And, given that the interval between the endpoints can pass the curve between two dashes, or, if the curve is twisted enough, the interval might pass it twice, the need of the concrete value is also questionable. The authors seem to ablate this by randomly permuting the auxiliary target of the training set, showing that the test accuracy stays on the level of the label-only task, but it doesn’t automatically prove that the concrete targets must be exactly δβ0\delta \beta_0.

W3. Is it topology that matters? An ablation I would love to see is, instead of using the proposed auxiliary loss, take a simple one, that would also be related to the query, e.g. the length of the line segment. It would allow one to decouple the potential topological explanation from the “query localisation”. Maybe all you need is just any target related to the query itself? My preliminary checks (seed 23) support this: I tested two possible auxiliary targets, the number of dashes in the intersection of a line segment between endpoints, and the distance between them. In both cases, the test accuracy was substantially higher, than for the label-only model (97.89%97.89\% and 91.11%91.11\% at the 200th epoch respectively). I believe such comparisons are essential for the experimental analysis, and I would love to see that kind of analysis in the paper.

W4. The auxiliary loss is not ablated. The motivation behind the auxiliary task is understood, the concrete implementation as the Huber loss, however, is not justified. Why do the authors prefer the Huber loss over MSE, L1, or a classification (since the target is just an integer in [0,12][0,12])? Authors don’t state the threshold δ\delta, used in the Huber loss (the code use δ=1\delta = 1), authors don’t ablate λ\lambda, using λ=0.5\lambda=0.5 for every Pathfinder run, and using λ=0.05\lambda=0.05 for DRIVE. I would love to see the sweep over λ\lambda in order to validate if the observed jump is robust or not.

W5. The paper needs a rewrite. The paper reads unclear sometimes, some parts of the texts are too repetitive, others - too vague. For example, authors compare the previous related work with their proposed approach in every paragraph of the corresponding section, before their approach is even described, making it hard to follow. The section of related work requires special proofreading: e.g. authors mention “Towards a general purpose CNN for long range dependencies in NND” by Romero et. al., saying that, together with other papers, this paper receives 1-dimensional pixel sequences. However that paper has 2D experiments as well (see, for example, Figure 3).

Questions for authors

Does paper really require topological treatment? The topology-inspired auxiliary target? The exact implementation of the loss as proposed? If the answer to any of the questions is “yes”, please explain why.

5 comments
AuthorsOpenReview ↗

Response to Reviewer BJBR

Thank you for independently reproducing the gain and testing alternative targets.

W1–W2 (role of topology). We agree that your component count and our proposed relative target coincide up to a constant: Δβ0=k−1\Delta\beta_0=k-1, where kk counts the foreground components met by the connector. Although this elementary count can be computed without homology, our formulation provides a principle for constructing supervision. The construction also gives conditions for invariance under data augmentation (§3.2). The same principle also generates the area-filtered degree-one road target (general response; SI §H).

W3 (simpler query-dependent targets). We tested both suggestions. Segment length is endpoint distance; we interpret the intersected-dash count as the number of foreground pixels within 1.5 pixels of the queried segment.

Auxiliary target4,000 scenes / 200 epochs2,000 scenes / 400 epochs
Segment length93.81 ± 5.4655.85 ± 0.36
Intersected-dash count98.37 ± 0.4557.74 ± 4.12
Relative Δβ0\Delta\beta_0 (proposed)98.52 ± 0.0697.11 ± 1.49

Validation-selected test accuracy (%), mean ± sample SD over three seeds; each target is divided by its training-set SD (this makes the new labels more stable for learning). Bold marks the highest mean, not significance. Budgets (epochs) approximately match gradient updates (SI §E.1).

Both alternatives help at 4,000 scenes as you expected. However, with a smaller data size at 2,000 scenes, the proposed relative target retains its benefit. Thus, we speculate that the topology also plays a role in addition to query localisation.

W4 (loss and weight). The exact Huber formulation is not required: L1 and integer cross-entropy also succeed. SI §E.2 specifies the default δ=1\delta=1 and reports the loss, δ\delta and λ\lambda sweeps; MSE and excessive auxiliary weight are less reliable.

W5 (presentation). We rewrote §2, explicitly separating two-dimensional continuous-kernel CNN experiments from one-dimensional sequence models.

AuthorsOpenReview ↗

Response to Reviewer BJBR: additional ablation on the auxiliary weight

Follow-up to W4 (auxiliary weight). As for the weight ablation, we tested on the standard Pathfinder setting (plain stack, 4,000 training scenes). Changing only the auxiliary weight from λ=0.5\lambda=0.5 to λ=0.05\lambda=0.05, the value used for DRIVE, gives 98.78 ± 0.22% validation-selected test accuracy (98.15 ± 0.39% at λ=0.5\lambda=0.5; +0.63 points; mean ± sample SD over three seeds), and all three seeds exceed 90% validation accuracy. Although it is not complete ablation study, this supports robustness to the choice of the auxiliary weight.

I thank the authors for the detailed response and the additional experiments. The experiments provide new information, so I will raise my Informativeness score from 1 to 2.

However, I’m not really convinced by the presented topological argument. The topological formulation as “a principle for constructing supervision” seems to me a weak justification for invoking the machinery of algebraic topology. In the plane, with a contractible AqA_q, the pair (X∪Aq,Aq)(X\cup A_q, A_q) has non-trivial homology only in degrees 0 and 1, so this “principle” boils down to counting either the components the segment meets or the loops it closes, and which of the two is used depends on the dataset and its labels (see Appendix H, “Choice of homological degree”).

As for W3, the target I suggested was not really tested. I asked for local intersections: the number of times the segment crosses ink, with repeated crossings of the same curve counted separately, i.e. the number of ink pieces along the segment, not the number of ink pixels near it. This pixel count lies somewhat in-between my target and the segment length, which may be why it turns out barely more informative than the length at 2,000 scenes. This 2,000-scene setting is the only evidence that the topological target beats non-topological ones: besides the 4,000- and 2,000-scene settings, the authors also report dense clutter (16,000 training scenes), and there, with identical scaling, the pixel count again matches Δβ0\Delta\beta_0 (Table 10). Unlike the pixel count, the piece count is an integer on the same scale as Δβ0\Delta \beta_0, so it can be compared without rescaling, like the unscaled Δβ0\Delta\beta_0 reference. The target I originally proposed might well match Δβ0\Delta\beta_0 at 2,000 scenes too. I would appreciate it if the authors added this experiment to Table 10.

Additionally, the segment-length control on Satellite PathX seems to be uninformative by design, since the query separation is restricted to 128±4128\pm 4 px (§5), and Appendix H acknowledges that “the segment-length target varies little because separation is restricted by the dataset design”. Predicting a near-constant carries no information, so its accuracy being close to label-only is expected, and it cannot show that “this non-topological query-dependent target does not provide the benefit”. The other controls are a scene-only count and a permuted target, so there is no informative non-topological control on this task at all.

More generally, for a method that is, in my view, weakly motivated, the evidence has to come from extensive ablations, and these are few. The loss, the weight and the Huber threshold are swept only for Δβ0\Delta\beta_0 (Appendix E.2), while the alternative targets in Table 10 are run in a single configuration. It might be that, with tuned hyperparameters, the pixel count would match Δβ0\Delta\beta_0. Given all this, I keep my rating.

Reviewer_BAsMOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose to address long-range dependency in query-dependent relational tasks in visual data using their developed concept of relative topology. Rather than modify the architecture to change how information propagates through the network, they introduce an auxiliary target during training. The paper emphasizes their zero-degree formulation which uses Betti-0 (connected components) in the auxiliary target. Claimed contributions are: 1) Formulation of the relative homology, 2) improving capability of feedforward CNNs and overall parameter efficiency for long-range dependency tasks without the need for substantial architectural changes and 3) demonstrations of the utility of the approach on real and synthetic datasets.

While the developed approach seems aligned and useful for the included datasets, the extension to query-dependent relational tasks in other modalities is not discussed-- the reviewer sees this as a significant drawback to the work since topological structures and properties are utilized in graph-based learning and many other problem spaces. The overall framing of the paper seems to suggest a desire to utilize topological features and properties rather than motivation to address a technological capability gap despite the title.

Critical strengths

The approach is successful on the datasets considered. Authors demonstrate improvement in training time (excluding calculation of the relative topology signal) and improved efficiency requiring fewer training scenes and model parameters than baselines to achieve 95% validation accuracy.

Critical weaknesses

Insufficient information regarding the computational components of the relative topology (tools being used to compute B0\Beta_{0}, time to calculate the topological targets, threshold being used and if it varies by image). The authors only report training time assuming that the topological targets they use for supervision are precomputed and available like their label counterparts. While a case could be made that we don't report labeling time for traditional supervised learning, these limitations should be transparently discussed. Overall, the broader context, trade-offs, and limitations should be presented for the readers. Empirical comparisons show marginal (even if statistically significant in some cases) gains. These gains are hard to contextualize without the full computational burdens being presented.

Literature review is missing citations to the broader field of topological regularization which has been used somewhat broadly in computer vision (TopoUT: Enhancing Cell Segmentation Through Efficient Topological Regularization, A Topological Regularization Term for Inverse Scattering Problems, On the Impact of Topological Regularization on Geometrical and Topological Alignment in Autoencoders: An Empirical Study, TopTemp: Parsing Precipitate Structure from Temper Topology, A Topological Regularizer for Classifiers via Persistent Homology, and others). While there may be differences in efficiency, scalability, and motivation the omission of these works left the reviewer wondering how the baselines were chosen for comparisons and why no other topological constraints or regularizers were considered in the experiments. They include the relative topology comparison to a global topology (again with limited computational implementation details).

Questions for authors

How extensible is the developed approach to other modalities and related problems? What would be the greatest barrier to generalizing the techniques?

Why are other topological penalties and loss regularizers insufficient baselines for the problem considered?

1 comment
AuthorsOpenReview ↗

Response to Reviewer BAsM

Thank you for raising comparison, computational cost and transfer.

Comparison and positioning. We agree that other topology-aware losses are relevant baselines. The added DRIVE comparisons cover centreline, persistent-homology and component-graph objectives, including the suggested TopoUT. The general-response table reports each objective alone and with our proposed relative supervision; four of five show a significant gain on at least one metric after Holm correction (SI §G.2). The revised Section 2 also includes the broader literature you identified.

Cost and implementation. One of the main advantages of our proposed relative supervision is it is computationally cheap: it is a one-time target construction at existing queries, with unchanged inference. For the DRIVE segmentation reference, training plus one-time target construction increases from 364.82 to 369.17 s: +4.35 s (+1.2%), with unchanged inference. This excludes query sampling and evaluation; SI §G.3 gives measurement conditions and costs for all evaluated objectives. Pathfinder uses SciPy connected-component labelling with 8-connectivity and a fixed foreground threshold of 0.1 for every image. SI §D.5 accounts for its target-construction cost; SI §E.4 shows that learning succeeds across the tested thresholds.

Transfer and its requirements. Satellite PathX applies the same principle in degree one to real-image road connectivity (general-response table; SI §H). Both real-image applications use existing structural annotations. Transfer to another modality requires a usable structural representation and a task-relevant query-defined subspace whose target is tractable to compute.

Reviewer_vLjCOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work proposes a supervised learning method for query-based visual tasks, using auxiliary labels derived from a query-dependent relative topological construction. The method is demonstrated on the Pathfinder same-path task, which asks whether two query positions lie on the same path. The auxiliary label, Δβ0\Delta\beta_0, is the change in the number of foreground connected components when the straight segment joining the two query positions is added, and Δβ0=1\Delta\beta_0 = 1 implies that the positions are connected. A network predicts this target alongside the same-path label, and the auxiliary head is removed at inference. The authors evaluate the method on a plain feedforward CNN against label-only training, a query-independent global topological target, and a permuted-target control, and also test residual and Swin-T backbones, compact models, limited training data, and retinal vessel segmentation on DRIVE.

Critical strengths
  • Model capacity and computational cost: The paper addresses both practical questions. It evaluates compact models as well as larger ones, and it reports the computational cost of the auxiliary target.
  • Application beyond the synthetic benchmark: The experiment on retinal vessel segmentation is an interesting application, as it tests the construction on real data, beyond the synthetic Pathfinder setting.
Critical weaknesses

Major

  • Limited methodological contribution: The only contribution I can identify is the extraction of the auxiliary label. This step is independent of the learning procedure itself: the models are standard architectures, trained in a standard supervised manner on the extracted labels. The paper should make clear what it contributes to learning beyond this label construction.

Minor

  • Backbone families not introduced: The model families used as backbones, such as Swin-T, are neither introduced nor referenced.
  • Panel references in Figure 2: The references to panels (a,b) and (a,c) appear to be incorrect, and panel (d) does not seem to be referenced at all.
Questions for authors
  1. Apart from the extraction of the labels, what is the methodological contribution of this work?
  2. The paper lists related work on the Pathfinder problem. How were the query points on the lines chosen in these works?
  3. In line 246, how is the function conv defined?
2 comments
AuthorsOpenReview ↗

Response to Reviewer vLjC

Thank you. We have clarified the methodological contribution and addressed the three questions (please see also the general response).

Q1 (contribution to learning). Retaining standard architectures and supervised training is a practical advantage: our proposed relative supervision enables a plain CNN to learn difficult long-range connectivity with the same backbone, optimiser and schedule (Table 1). Our contribution is the principle for constructing that signal. A global invariant stays constant across queries on a fixed scene; our construction makes the target describe how the query-defined subspace interacts with the scene (§3). The general-response tables demonstrate another target (constructed with the same proposed principle) on Satellite PathX and additional benefit alongside existing DRIVE objectives, with unchanged inference (SI §§G–H).

Q2 (query placement in prior work). Linsley et al. (2018) mark target-path endpoints: both ends of one path for positives, one endpoint from each of two paths for negatives; distractors are not queried. Kim et al. (2020) use that generator, and our endpoint placement follows the same rule (SI §B).

Q3 (definition of conv). conv⁡\operatorname{conv} denotes the convex hull, hence the closed straight segment for two points; this is now explicit in §3.2.

Minor presentation points. We added backbone descriptions and citations (§4; SI §C) and corrected Figure 2's panel references, including the dense-clutter panel (d).

I thank the authors for their clarification. However, I remain on the point that the methodological contribution beyond the label construction remains limited. I will therefore keep my scores.

Reviewer_Zu1GOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes using relative topology as an additional supervisory signal to help neural networks learn long-range relationships. The authors argue that global topological features do not necessarily capture the relationship between two selected points. To address this, they connect the two points with a temporary straight segment and calculate how the number of connected components changes. This quantity is used as an additional training target. The authors evaluate their approach on Pathfinder, where the goal is to determine whether two points belong to the same path, and retinal vessel segmentation. They show that this additional supervision can significantly improve the performance of ordinary CNNs on Pathfinder, while also improving some structural metrics on retinal vessel segmentation.

Critical strengths
  • The paper presents an interesting idea. Instead of designing a more complicated neural network to capture long-range dependencies, the authors try to improve the supervision. I also find it interesting that the additional target can be computed from the available training data without requiring extra annotations, and the auxiliary prediction head is removed during inference.
  • The results on Pathfinder are very good. They also include useful control experiments, such as global topology, permuted targets, and self-pretraining, which help to show the importance of the proposed supervision. The additional experiments with smaller networks and retinal vessel segmentation are also useful.
Critical weaknesses
  • The main limitation is that most of the experimental evidence comes from a synthetic dataset. Although the results on Pathfinder are impressive, it is unclear whether the method generalizes to more complicated real-world problems. The authors also evaluate retinal vessel segmentation, but they use only DRIVE, with 16 training images and 20 test images. Also, improvements in conventional segmentation metrics are relatively small. For example, Dice improvements is not statistically significant. The improvements in structural metrics are more convincing, but I think at least one additional real-world dataset would strengthen the conclusions.
  • Another concern is the lack of comparison with existing topology-aware losses. The authors discuss methods such as clDice and persistent homology based losses, but they do not compare against them in the retinal segmentation experiments. Although the control experiments are useful, it is unclear whether the proposed supervision performs better than existing methods designed to preserve topology. Also, the paper only investigates one way of constructing the relative topological target, using a straight connector between two points. I did not understand how sensitive the method is to this choice, or whether other constructions would provide similar results.
Questions for authors
  • Can the authors compare their method with existing topology-aware losses, such as clDice or persistent-homology-based losses, on retinal vessel segmentation? How would the results change if the authors used a different reference subspace? For example, what happens if the width or shape of the connector changes?
  • Can the authors provide experiments on another real-world dataset? It would be beneficial to understand whether the large improvement on Pathfinder can generalize to more complicated tasks.
1 comment
AuthorsOpenReview ↗

Response to Reviewer Zu1G

Thank you for the three concrete suggestions.

Additional real-image evidence. Satellite PathX tests road connectivity in satellite imagery using an area-filtered degree-one target from the same principle. Our proposed relative target reaches 79.38 ± 1.07% and exceeds all four controls (Holm-adjusted p<0.001p<0.001) (general-response table; SI §H).

Comparison with topology-aware losses. The general-response DRIVE table compares five existing objectives, including clDice and persistent-homology losses, both alone and with our proposed relative supervision. Four of five show a significant gain on at least one metric after Holm correction; SI §G.2 reports the full results for each objective and metric.

Connector width and shape. On Pathfinder, all tested capsule radii (0.5–4.0 pixels) reach 90% validation accuracy in 3/3 seeds. For both filled-ellipse variants, the success count rises from 0/3 with unscaled targets to 3/3 after scaling by the training-set SD. Thus the construction supports several widths and shapes, with target scale affecting learning (SI §E.3).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers, including Reviewer BJBR for independently reproducing our main result. The revision addresses the main concerns as follows:

  • Comparison and positioning: we compare existing topology-aware objectives with and without our proposed relative supervision on DRIVE.
  • Generality beyond synthetic data and degree zero: we add Satellite PathX on SpaceNet 3, using a degree-one target from the same principle.
  • Computational burden: we measure training and target-construction costs; inference is unchanged.

First, Satellite PathX is a real-image analogue of synthetic Pathfinder: it asks whether a road route connects two marked locations within a satellite image. We construct its (area-filtered) degree-one loop-change target (Δβ1\Delta \beta_1) using the same query-defined principle (SI §H).

MetricLabel-onlyGlobal loop countPermuted targetSegment lengthRelative target (proposed)
Test accuracy (%) ↑50.07 ± 0.1950.05 ± 0.3656.01 ± 1.2251.40 ± 1.8079.38 ± 1.07 ↑***

Our proposed relative target improves accuracy over every control (Holm-adjusted p<0.001p<0.001, paired tile-level tests; SI §H). Tables report mean ± sample SD over three seeds at validation-selected checkpoints.

Second, on DRIVE we compare established topology-aware objectives, which act on segmentation outputs, with and without our proposed relative supervision. An auxiliary head predicts our target from shared features, while the backbone, optimiser and schedule remain fixed.

Training objectiveDice (%) ↑ (base → + ours)clDice (%) ↑ (base → + ours)
BCE + Dice (reference)79.95 ± 0.23 → 80.15 ± 0.23 ↑80.97 ± 0.10 → 81.67 ± 0.37 ↑
clDice (CVPR 2021)79.19 ± 0.15 → 79.22 ± 0.07 ↑83.09 ± 0.11 → 83.35 ± 0.32 ↑
TopoLoss (NeurIPS 2019)80.06 ± 0.16 → 80.35 ± 0.07 ↑*80.89 ± 0.49 → 81.71 ± 0.55 ↑***
Betti matching (ICML 2023)61.36 ± 1.05 → 61.74 ± 0.39 ↑74.28 ± 1.00 → 73.07 ± 3.03 ↓***
Topograph (ICLR 2025)79.55 ± 0.16 → 79.81 ± 0.04 ↑**80.41 ± 0.31 → 81.01 ± 0.49 ↑***
TopoUT (ISBI 2024)77.42 ± 0.24 → 77.90 ± 0.12 ↑79.21 ± 0.17 → 81.05 ± 0.28 ↑***

Entries show without → with our proposed relative supervision (weight 0.05). Bold marks the best mean. */**/*** denote Holm-adjusted p<0.05/0.01/0.001p<0.05/0.01/0.001 for the difference between base and ours; ↑/↓ indicate the direction of changes.

In the Dice/clDice table, adding our proposed relative supervision increases 11 of the 12 mean scores; only clDice for Betti matching decreases. Note that Dice/clDice are the most widely reported metrics for DRIVE. Full comparisons including other metrics are in SI §§F–G.

Both applications use existing training annotations (no extra data or annotations). For the DRIVE segmentation reference, overall training time (training plus one-time target construction) increases from 364.82 to 369.17 s, an additional 4.35 s (1.2%; SI §G.3).

We'd like to emphasise that two-dimensional Pathfinder is a longstanding challenge for learning long-range connectivity (§2). Our proposed relative supervision enables a plain CNN without residual connections, recurrence or attention to learn this task by changing the training signal alone: accuracy rises from 61.30% to 98.15% with 4,000 training scenes and the same backbone, optimiser and schedule (Table 1). This auxiliary-target formulation also makes it straightforward to add to existing objectives, as the DRIVE comparisons demonstrate, with unchanged inference.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.