acceptodds
Withdrawn from ICLR 2027

Learning Task-Conditioned Utilities for Topological Errors in Thin-Structure Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Topological errors with similar overlap can have different functional consequences, and the same gap, spur, or bridge can have different utility across tasks. Existing topology-aware losses encode structure but usually fix the relative cost of error types. We formulate topology supervision as learning a task-conditioned utility over a finite, interpretable edit signature, with held-out functional outcomes separated from the segmentation-training objective. TOPOUTILITY represents gaps, false bridges, spurs, loops, endpoint displacement, and branch mismatch, and defines diagnostics for utility identifiability, ranking stability, grouping stability, and extractor stability before weights are frozen or cross-fitted. On DRIVE, it reaches 0.86 pairwise agreement, 0.83 severity correlation, 7.2% downstream path failure, and 0.858 clDice at 21.5 ms; fixed graph-edit costs reach 0.78, 0.74, 10.9%, and 0.847 at 36.7 ms. These reported aggregates validate held-out predictive utility and its downstream association; the full identification diagnostics remain separate evidence requirements.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
topology-aware segmentation, thin-structure segmentation, task-conditioned utility, structural errors, topology-aware learning, image segmentation, functional utility, connectivity-aware learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_cRndOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper argues that topological errors in thin-structure segmentation should be judged by their functional consequence for a specific task, not by pixel overlap or by a fixed topology loss. It defines a six-dimensional interpretable edit signature, learns task-conditioned weights over these signatures from held-out preferences or functional outcomes, and then freezes or cross-fits the learned utility into segmentation training. The central methodological point is a two-stage protocol that keeps utility estimation separate from segmentation training, so that segmentation gains cannot be used to validate the utility itself.

Critical strengths
  1. The problem framing is clear and relevant: overlap metrics and fixed topology losses do conflate error detection with task-dependent consequence.
  2. The two-stage protocol directly addresses the circular-validation risk, which is a real weakness of “learn the loss from the final segmentation result” approaches.
Critical weaknesses
  1. Core evidence is missing. Rank/conditioning, bootstrap intervals, matched-overlap AUROC, identical-ordering, and collinearity results are not shown. These are precisely the evidence needed for H1/H4, so the paper currently supports predictive association, not identification.
  2. Only one dataset is reported. DRIVE is the only results table; the crack benchmark appears only in the protocol. This is a serious gap for a paper whose main claim is task-conditioned utility: without at least two tasks with different orderings, the "task-conditioned" part is not directly demonstrated.
  3. No multi-task or cross-task evidence. The paper never shows two tasks producing distinct utility orderings, nor the killer experiment with identical true orderings. The central scientific claim is therefore asserted rather than tested.
  4. Limited technical novelty. The method is a linear softmax weighting over six hand-defined features plus a Bradley–Terry likelihood.
  5. Insufficient implementation detail. The task descriptor q, the feature map ϕ, the extractor algorithm and its hyperparameters, and the preference-collection procedure are not specified, which makes the method hard to reproduce or assess.
  6. Statistical reporting is thin. The protocol promises at least five seeds and bootstrap intervals, but the table reports only point estimates; there are no confidence intervals, effect sizes, or multiple-comparison corrections.
Questions for authors
  1. The paper states that rank and conditioning, bootstrap-interval, matched-overlap AUROC, identical-ordering, and collinearity results are required but not displayed. Can you provide these results, and do they support H1 and H4?
  2. What exactly is the task descriptor q, and what does ϕ(q) encode? Is it a one-hot task label, hand-crafted features, or a learned representation?
  3. How are the six edit signatures extracted? Which skeleton/correspondence algorithm and hyperparameters are used, and how sensitive are the results to this choice?
  4. What are the crack-benchmark results? If they are not available, why does the protocol declare them?
  5. How many preference labels and annotators were used, and what is the inter-annotator agreement?
  6. Why is only DRIVE reported when the protocol declares two benchmark families?
Reviewer_7X4wOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes

Critical strengths

This is an important problem. Maintaining the connectivity of thin connected structures is important. It is also true that, depending on the downstream task, different topological issues may be of different importance.

Critical weaknesses

Prior art is 7 lines, one pargraph. Two strings of citations. This does not seem like due diligence. This paper is an abstract, it is not complete enough for a reader to understand or properly evaluate the work.

This seems to be an idea that will lead to an interesting paper --the fact that a table with numbers exists indicates that something may have actually been done. Unfortunately, I am unsure of what it is. I am missing about 1/2 page of prior art and maybe 5 pages of exposition. This paper is more of an outline.

For a paper that aims to defend the points made in this one, I would expect an explanation of what the specific downstream tasks are. I would expect to see some model trained with this loss (an explanation of which model) and how the resulting segmentation improved. I would expect to see images of the resulting segmentations for those tasks with insets or arrows indicating where the loss improved the results.

Questions for authors

What model is "fit"?

How do these numbers compare to SOTA on DRIVE or crack segmentation?

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

The paper is set out to evaluate (and possibly improve, I'm unsure) segmentations based on their functional downstream utility. It defines a set of error types and learns their respective weight (i.e., importance) based on the performance in the downstream application. The method is evaluated on one dataset.

Critical strengths

Evaluating the downstream utility directly is theoretically preferable over evaluating a proxy, such as pixel-wise or topological accuracy.

Critical weaknesses

The quality of the paper is extremely low, such that a detailed list of all the weaknesses is not feasible. The main weaknesses leading to this low quality are that (1) the writing is poor, (2) the paper is not embedded in existing literature, (3) the method itself is questionable, it is not clearly described, and the evaluation is unclear, and (4) the paper makes claims that it does not hold up.

(1) Writing. It took me multiple reads of the entire paper to somewhat understand what is going on. The paper does not follow a clear storyline. It uses terms without defining or describing them, or that are unclear in the context they are being used. In general, it is extremely hard to follow the paper. In fact, I still do not fully understand what paragraphs 2 and 3 of the intro are trying to tell the reader. Examples of terms that are/were unclear to me in the context they appear in include (non-exhaustive): "edit signature", "error signature", "utility identifiability", "grouping stability", "extractor stability", "path failure", "flow failure", "reportability test for task-specific edit costs", "trainable graph distance", "task descriptor".

(2) Related work. It basically does not exist.

(3) Method and evaluation of it. First, I am questioning whether a method that requires not only segmentation labels but also "preference" labels (did I understand this right?) is even desirable (because it removes any kind of method generalizability) and achievable (because I think in most cases it is very, very hard to define which errors are better or worse than others). Additionally to this question, I think the method is poorly described. I still don't understand how the supervision for "application utility" is obtained, e.g. for the drive dataset. I don't understand the evaluation entirely: some baselines and some metrics are not described, some others are being left out, etc.

(4) Claims that are not being held up. For example, the second dataset or the four hypotheses (at least not explicitly).

Questions for authors
  • What is P0, and P1?
  • What are procedural branching graphs?
  • What is the "killer experiment"?
  • Where are CrackTree results?
  • What is "Fixed GED"?
  • How does this work in detail: "Path connectivity or flow simulation supplies objective functional severity."

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

Related papers

Discussion (0)

Sign in to comment.