acceptodds
Under review as a conference paper at ICLR 2027

Cross-view Instance Assembly in Natural-color Tomography

Abstract

Existing segment-everything methods were built for video, three-dimensional scenes, grayscale CT data, and artificially colored microscopy. Natural-color tomography, which reveals the interior of opaque materials in RGB, is not among these domains. How well those methods transfer to this domain is largely unknown. We propose WINZE, a class-agnostic instance segmentation method designed for natural-color volumes. WINZE reads the volume section by section in three orthogonal orientations with a frozen SAM 2 backbone and assembles the sections into volumetric instances by rule without training. We compared it with seven segment-everything methods on 1,239 manually annotated instances. WINZE ranks first in instance [email protected] in four of the six scored scopes and achieves scores 1.3 to 3.1 times those of the best compared method on the three sets of scored planes. Our results give a reference point for this domain, motivate methods that build the spatial continuity of objects into their assumptions, and suggest that the main difficulty lies in holding each object as one label across sections.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
volumetric instance segmentation, cross-view consistency, foundation-model transfer, sparse annotation, topology

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LFHfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces WINZE, a class-agnostic and training-free approach for volumetric instance segmentation in natural-color tomography. According to the supplied submission description, WINZE applies a frozen SAM 2 model section-by-section in three orthogonal orientations and uses rule-based cross-view assembly to convert the resulting 2D masks into volumetric instances. The reported evaluation comprises 1,239 manually annotated instances and seven segment-everything baselines. The abstract claims that WINZE ranks first in instance [email protected] on four of six scored scopes and reaches 1.3–3.1× the strongest comparator on three sets of scored planes. The problem is practically relevant: generic 2D foundation models do not directly provide stable volumetric identities, and dense 3D annotation is expensive.

Critical strengths
  1. Practically consequential problem formulation

Leveraging a frozen 2D foundation model for volumetric instance segmentation without task-specific training is attractive where dense 3D annotation is costly. The proposed use of three physical views directly targets the known limitation that independent slice-wise processing discards volumetric structure. Prior literature explicitly distinguishes slice-wise approaches from volume-wise approaches and notes that independent 2D segmentation does not use 3D structure.

  1. Potentially useful annotation and benchmarking effort

A manually annotated set of 1,239 instances, if independently sampled and carefully quality-controlled, could be a meaningful resource in a data-scarce application. The comparison with seven segment-everything approaches is broader than a single-baseline demonstration and may reveal a real practical advantage.

Critical weaknesses
  1. The stated evaluation does not clearly validate the central 3D claim

The decisive concern is the apparent mismatch between volumetric instance assembly and scores reported on planes/scopes using instance [email protected]. Slice-level agreement can remain high even when a method: merges two touching objects between annotated planes; fragments one object across depth; switches identity after an occlusion or narrow neck; invents or removes handles, cavities, branches, or disconnected components; produces mutually inconsistent labels across orthogonal views. These are not corner cases; they are the defining errors of volumetric instance assembly. By comparison, prior slice-to-volume work motivates propagation specifically from cross-slice consistency and evaluates the association component against a no-association slice baseline.

  1. Novelty and causal contribution are insufficiently isolated

Applying a frozen segmenter to orthogonal slices and reconciling masks is intuitively reasonable, but the contribution may reduce to a collection of task-specific rules unless the paper demonstrates that the assembly procedure embodies a general and non-obvious principle. The relevant literature already includes orthogonal annotation, orthogonal ensembles, multi-view co-training, multi-view fusion, masked multi-view volumetric segmentation, and learned slice-to-volume propagation; representative bibliographic entries include Orthogonal Annotation (CVPR 2023) [1], Orthogonal Ensemble Networks (MICCAI 2021) [2], S2VNet (CVPR 2024) [3], and SwinMM (MICCAI 2023) [4].

[1] Orthogonal Annotation Benefits Barely-Supervised Medical Image Segmentation. CVPR 2023.

[2] Orthogonal Ensemble Networks for Biomedical Image Segmentation. MICCAI 2021.

[3] Clustering Propagation for Universal Medical Image Segmentation (S2VNet). CVPR 2024.

[4] SwinMM: Masked Multi-view with Swin Transformers for 3D Medical Image Segmentation. MICCAI 2023.

Questions for authors
  1. What genuinely volumetric ground truth and metrics are available? Please report, at minimum, full-volume instance precision/recall/F1 at multiple 3D IoU thresholds, panoptic quality or AP, split and merge counts, fragmentation per ground-truth object, identity consistency across depth, and connected-component/topology errors. If only sparse planes are annotated, please explain exactly which 3D claims can be inferred and provide a validation study comparing sparse-plane ranking with densely annotated subvolumes.

  2. How were the 1,239 instances sampled and annotated? Please report the number of independent volumes, specimens, materials, scanners, acquisition settings, annotators, adjudication protocol, inter-annotator agreement, object-size distribution, and whether multiple planes from one object are counted as independent observations. What is the statistical unit used for uncertainty estimates?

  3. Were any annotations used to design or tune WINZE? For every threshold and rule, identify its value, search range, tuning specimens, and whether those specimens overlap with reported evaluation volumes. Please provide leave-one-specimen-out or held-out sensitivity results.

  4. What is the exact assembly algorithm? Please provide pseudocode defining mask generation, within-view linking, cross-view matching, graph construction, transitive closure, voxel assignment, unmatched masks, conflicting labels, merge/split resolution, nested objects, holes, disconnected components, and the order in which rules are applied.

  5. Can the authors isolate the source of improvement? Please provide one-/two-/three-view results; leave-one-orientation-out results; assembly-rule deletions; no-assembly and simple-fusion baselines; SAM versus SAM 2 under the same pipeline; and backbone-only versus filtering-only versus assembly-only contributions.

  6. How does the method behave outside the reported acquisition regime? Please stratify results by voxel anisotropy, resolution, object density, object diameter, elongation, branching, contact/overlap, color contrast, partial-volume artifacts, and view-dependent appearance.

  7. What are runtime, peak memory, and storage costs per volume? Three-orientation sectioning can multiply SAM 2 inference and mask-storage costs. Please report hardware, number of sections and masks, wall-clock time by stage, scaling with volume dimensions/object count, determinism, and compute-matched baseline results.

2 comments
AuthorsOpenReview ↗

Response to Reviewer LFHf (1/2)

We thank the reviewer for highlighting the distinction between agreement on individual planes and maintaining consistent object identities in 3D. Following the reviewer’s suggestions, we conducted additional experiments mainly on the 3D evaluation. They showed that the plane-level evaluation is consistent with the 3D evaluation, and clarified the kinds of 3D errors that WINZE makes. We respond to each point below.

W1: Validating the 3D assembly beyond plane-level F1

Our evaluation does not rely solely on agreement within the XY planes. The scored XZ and YZ planes can reveal splits and merges across depth where those errors intersect the annotated planes. The densely annotated 3D ROIs provide a direct 3D evaluation (Section 4.3). Following the reviewer’s comment, we compared F1 results in the planes and in the 3D ROIs for all pairs of the eight methods using three volumes (revised Table 2). In ~93% of the pairs, the same method showed higher scores. The two evaluations thus gave consistent results on the volumes we evaluated (Appendix G.5).

We also counted the split and merged objects in 3D, which plane scores cannot fully capture. We compared WINZE with three alternatives: no whole-mask grouping, IoU linking, and independent IDs per XY section. WINZE still produces fragmented instances, but its fragmentation per ground-truth object was equal to or lower than in all three (Table 26). The no-association slice baseline had no true positives in the 3D ROIs, indicating that association across sections is necessary (Tables 23 and 26). In the revised manuscript, we present these results and clarify the role and scope of each evaluation (Section 4 and Appendix G.5).

W2: Contribution relative to prior orthogonal and multi-view methods

The novelty of WINZE lies not in using orthogonal slices but in the principle of determining 3D labels without altering the shape of the 2D masks of SAM 2. Because of this principle, WINZE is expected to apply to a broad range of volume data that SAM 2 can segment (Section 6). The ablation study isolated the contribution of this mechanism, as removing whole-mask grouping or track filtering lowered F1 substantially (Table 23). Removing the stages taken from u-Segment3D changed F1 little (Section 5.3). WINZE is therefore not merely a collection of task-specific rules.

The four cited methods assume training with annotations and output class-specific regions. WINZE differs from them in that it produces instance labels without additional training. We state this in the revised manuscript (Appendix A).

Q1: Volumetric ground truth and 3D metrics

Following the reviewer’s suggestion, we report instance precision, recall and F1 at 3D IoU thresholds of 0.3 and 0.5, split and merge counts, panoptic quality, fragmentation per ground-truth object, and connected-component counts in three densely annotated 64³ ROIs (Tables 22, 23 and 26). No full-volume ground truth is available because ground-truth data for natural-color tomography remain scarce. We thus used sparsely annotated planes instead. On the evaluated cubes, the order of F1 scores among the compared methods was broadly consistent between the annotated planes and the densely annotated subvolumes (Table 26). The plane-level evaluation therefore reflects the relative performance of methods in 3D. We state this in the revised manuscript (Appendix G.5).

Q2: Sampling, annotation protocol and statistical unit

Following the reviewer’s suggestion, we expanded the information on the annotation protocol and specimens. The submitted manuscript already included the numbers of volumes and specimens, the imaging methods and voxel dimensions, and the selection of cubes, planes and ROIs. The annotation procedure, which used one annotator without measuring inter-annotator agreement, was also described in it (Section 2.2 and Appendix B). In the revised manuscript, we additionally reported the size distribution of the ground-truth objects (Table 31). We also stated that there was no adjudication, that multiple sections of an object are not counted as independent observations, and that the specimen is the statistical unit (Section 4.3 and Appendix B.1). We did not report interval estimates because only two specimens for evaluation were available, which is too few for a meaningful estimate (Section 4.3).

AuthorsOpenReview ↗

Response to Reviewer LFHf (2/2)

Q3: Use of annotations for tuning and held-out sensitivity

Ground-truth annotations were not used to design WINZE or to set its thresholds. The thresholds were selected through qualitative inspection of the outputs on the development cube rather than a numerical search. Some thresholds were also set on a cube from the same volume as DS00136_cube0004 that does not overlap it (Appendix B.1). There is thus no search range. For each threshold and rule, the revised Appendix C identifies its value and any overlap with the reported evaluation volumes. As suggested, we performed a held-out sensitivity analysis on the evaluation cubes by varying each main threshold by ±20%. F1 changed little for most thresholds (Table 25).

Q4: Exact assembly algorithm

We provide pseudocode for the assembly procedure as Algorithm 1 in Appendix C, covering the items the reviewer lists. WINZE has no splitting stage and no constraints that guarantee the preservation of nested structures, holes or connectivity. These are thus not included in the pseudocode.

Q5: View, rule and baseline comparisons

We provided the suggested results to isolate the source of improvements (Tables 23, 24 and 29). F1 dropped substantially when removing whole-mask grouping, which uses both XZ and YZ, or track filtering (Table 23). These stages newly added in WINZE therefore explain the improvements (Section 5.3). For SAM versus SAM 2, using SAM (ViT-B) reduced F1 in almost all comparisons (Table 29).

Q6: Behavior outside the reported acquisition regime

Following the reviewer’s suggestion, we simulated changes in acquisition conditions by subsampling along Z, lowering in-plane resolution, reducing contrast, adding noise and blurring the existing data (Table 30). Performance decreased when the in-plane resolution was lowered, which shows that in-plane resolution is important when applying WINZE (Appendix G.9). We addressed the requested stratifications except for view-dependent appearance, which cannot be generated by modifying the data (Appendix G.7). The stratified analyses showed that WINZE was more likely to miss small objects and objects with low color contrast against the background. These are, however, possibly due to limitations of SAM 2, not those of newly developed parts. In contrast, touching objects were detected at least as well as isolated objects (Tables 11, 15 and 28).

As natural-color tomography is a recent technique, additional high-quality ground-truth datasets beyond those reported here are currently unavailable.

Q7: Runtime, memory and storage

We measured the requested quantities and report them in Tables 19 and 27. As a summary, hardware and wall-clock time by stage were already reported in Appendix F and Table 19 of the submitted manuscript. We additionally measured the numbers of sections and masks, peak memory, storage, scaling with volume dimensions, and determinism (Table 19). For scaling with object count, we report detection and assembly times as a function of mask count, which serves as a proxy for object count (Table 19).

Compared with the seven existing methods discussed in the manuscript, WINZE required a longer runtime, approximately 14 times their mean runtime (Table 27). Its peak GPU memory use (9.6 GiB) was within the range of the existing GPU-based methods (2.4–14.0 GiB). Storage per volume, including intermediate files, was approximately 1.3 GB for WINZE, compared with 0.08–351 MB for the existing methods (Table 27). These results show that WINZE can improve the results with reasonable resource requirements.

Reviewer_Y3djOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies 3D instance segmentation for natural-color tomographic volumes. The authors propose WINZE, a training-free pipeline that applies a frozen SAM2 model to sections from three orthogonal orientations and then assembles the resulting 2D masks into volumetric instances using rule-based reconciliation, tracking, filtering, and whole-mask grouping. The method is evaluated against seven existing segment-everything approaches using three exhaustively annotated cubes, one positive-only qualitative cube, and eleven public volumes containing one labeled object each. WINZE scores highest in most scopes against seven compared methods.

Critical strengths

(1) The paper addresses an underexplored and practically relevant problem. It provides useful empirical evidence on how existing segmentation methods transfer to natural-color tomographic volumes.

(2) The evaluation is transparent and reasonably broad. The paper distinguishes development from independent evaluation data, explains how evaluation regions were selected, and compares methods from several relevant segmentation families.

(3) The analysis of limitations and failure modes is informative. The authors report recall, precision, foreground coverage, and label counts, distinguish different types of failure, and openly acknowledge several limitations of the experiments.

Critical weaknesses

(1) The methodological contribution is limited. The first claimed contribution is primarily an empirical comparison rather than a technical advance. In the second claimed contribution, the major feature of WINZE (segmenting sections along three orthogonal orientations and assembling them) is not novel, as this strategy has already been established by u-Segment3D. In the fourth claimed contribution, describing the reported results as a "reference point" appears overstated, given the limited dataset scale and the absence of a broadly reusable benchmark. Consequently, it remains unclear whether the paper provides a sufficiently distinct and generalizable methodological contribution beyond a domain-specific engineering pipeline.

(2) The ablation study is weak. Across the entire paper, only Tables 18 (RGB vs. grayscale), 20 (different PPS values), 16 (full WINZE vs. simple IoU linking), and 14 (grouping-stage analysis) can be regarded as ablations. Even these experiments do not isolate the many key steps of WINZE in Figure3 or establish their individual effectiveness.

(3) The presentation of the evaluation scale is potentially misleading. The reported 1,239 annotated instances combine 1,045 two-dimensional instances from nine planes, 130 three-dimensional ROI objects, 53 positive-only qualitative objects, and one labeled object from each of 11 public volumes. These annotations are heterogeneous and should not be interpreted as 1,239 independent volumetric test instances.

(4) The method is poorly illustrated due to the lack of detailed diagrams. Figure 3 is the only central method overview, and it merely lists the stage names. It does not visualize the potentially novel mechanisms: atom formation, cross-section identity assignment, boundary-preserving tracking, track filtering/restoration, or whole-mask hyperedge grouping. Instead, these mechanisms are distributed across dense prose and appendix-level rules with code-like names. Consequently, it is difficult to understand how information flows through WINZE, to distinguish its novel components from those inherited from u-Segment3D, or to assess why the proposed rules should work.

Questions for authors

Regarding the critical weaknesses above, I would like the authors to address the following concerns.

(1) Please clarify the paper’s key contributions and specify where its methodological novelty lies.

(2) Please explain how the existing ablation experiments demonstrate the effectiveness of WINZE’s core components.

(3) Please justify how the experimental sample size is counted and reported.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Y3dj

We thank the reviewer for appreciating our evaluation of how existing methods transfer to this domain and our analysis of failure modes. Following the reviewer’s suggestions, we conducted additional experiments and revised the description of our contributions and the presentation of the results. These changes clarify the novelty of WINZE, the contribution of each stage, and the scale of the evaluation. We respond to each point below.

W1: Empirical contribution, orthogonal assembly from u-Segment3D, and the “reference point” claim

For the first contribution, we integrated it with the fourth contribution as a key comparison to highlight a general problem in the segmentation task on natural-color tomography.

For the second contribution, the novelty of WINZE lies not in the framework of assembling sections from three orthogonal orientations but in tracking XY instance masks without altering their shape. To avoid misreading, we replaced “in three orthogonal orientations and assembles the sections” with “without altering their shape and uses the orthogonal sections” (Section 1). With these changes, we have clarified the novelty of WINZE.

For the fourth contribution, we rephrased the sentence, removing “a reference point” to avoid overstatement (Section 1).

W2: Isolating the key steps of WINZE

Following the suggestion, we conducted stage-removal ablations that evaluate the stages of WINZE individually (Tables 23 and 24). Removing whole-mask grouping or track filtering substantially reduced F1. Our results show that these newly added mechanisms improved performance (Section 5.3).

W3: Heterogeneous annotations in the 1,239 instances

Following the suggestion, we reported the four types of annotations, previously integrated as 1,239 annotations, separately in the revised manuscript (Sections 1, 2.2 and 4.3). The 1,239 annotations can no longer be misread as independent volumetric test instances.

W4: Illustration of the mechanisms in Fig. 3

Following the suggestion, we revised Fig. 3 to distinguish the stages newly added in WINZE from those inherited from u-Segment3D (Fig. 3). The main text now names the stages by their function, and a table in Appendix C.1 gives their names in the source code. We also added ablations to evaluate how individual rules affect performance and stated the resulting improvement in the revised manuscript (Section 5.3 and Appendix G.2).

Q1: Methodological novelty

Following the suggestion, we stated the methodological novelty of WINZE in the revised manuscript. WINZE tracks XY instance masks without altering their shape and uses the orthogonal sections as evidence to judge the continuity of objects, whereas previous methods do not keep the shape of the SAM 2 masks (Section 3.1). The ablation study showed that the results of WINZE depend on this mechanism, not on the stages taken from u-Segment3D (Section 5.3 and Table 23).

Q2: Evidence that the core components work

In the revised manuscript, we added ablation experiments that show the effectiveness of WINZE’s core components (Appendix G.2). Removing whole-mask grouping or track filtering substantially reduced F1, showing that both newly added processes contribute to performance (Section 5.3).

Q3: How the sample size is counted

We reported the four types of annotations, previously integrated as 1,239 annotations, separately in the revised manuscript (Sections 1, 2.2 and 4.3). The four types are as follows: 1,045 2D instance labels on nine planes of the three annotated cubes; 130 3D objects in three 64³ ROIs; 53 objects in the qualitative cube; and one object in each of the 11 public volumes. In the revised manuscript, we replaced descriptions based on the total with explicit annotation types and counts.

Reviewer_jT4BOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a method which is training-free, class-agnostic instance segmentation framework for natural-color tomographic volumes. The method uses a frozen SAM backbone to generate 2D segmentation masks across three orthogonal views, followed by rule-based cross-view fusion, tracking, and grouping to reconstruct consistent 3D instances. The method is evaluated against seven existing segmentation approaches on natural-color tomography data.

Critical strengths
  • The paper addresses a relatively unexplored problem of instance segmentation in natural-color tomography, where existing segmentation methods struggle to generalize.

  • WINZE outperforms competing methods in four of six quantitative evaluation scopes (Table 2). It also successfully recovers 40 of 53 annotated objects in the qualitative cube, compared to only 6 for DEVA (Table 4).

Critical weaknesses
  • Novelty: The proposed approach primarily combines an existing SAM 2 backbone with several handcrafted rules for cross-view fusion, tracking, and instance grouping. Some components are also adapted from u-Segment3D (Section 3.1). Consequently, the methodological contribution beyond integrating existing techniques is unclear.

  • Ablations: Despite the large number of processing stages, the paper does not provide quantitative stage-removal ablations. In particular, the contribution of multi-view fusion and whole-mask grouping remains unclear. Although Table 16 provides a SAM 2 overlap-linking baseline, its detection settings differ from those used in WINZE, making it difficult to isolate the improvements from instance assembly.

Limited Evaluation and Generalization: The primary quantitative evaluation relies on only three volumes, with annotations restricted to selected planes and small 3D regions. Moreover, the absolute instance F1 scores remain relatively low (0.128–0.258 across the three-plane evaluations in Table 2). Orthogonal-view precision is particularly poor (Table 8), suggesting that consistent 3D instance assembly remains unresolved.

Questions for authors
  • The authors should improve the presentation in the paper. One way could be to add visuals for e.g. in the method pipelne figures Fig. 3. Also some places has two many specific implementation details which are generally avoided in the main draft for e.g. mentioning filenames in Line 163.
1 comment
AuthorsOpenReview ↗

Response to Reviewer jT4B

We thank the reviewer for recognizing the challenges of natural-color tomography and for the valuable comments on our comparisons and method description. Following the reviewer’s suggestions, we ran additional experiments and clarified the contribution of the mechanism specific to WINZE. We respond to each point below.

W1: Contribution beyond combining SAM 2 and u-Segment3D

The distinctive feature of WINZE is that it tracks the XY instance masks generated by SAM 2 without altering their shape and assembles them into 3D labels, whereas previous methods do not keep the shape of the SAM 2 masks. The ablation study also shows that WINZE owes its improved scores to this mechanism, not to the stages taken from u-Segment3D (Section 5.3 and Table 23). In the revised manuscript, we distinguish WINZE’s tracking and assembly procedure from the stages inherited from u-Segment3D and state that this mechanism was newly developed in WINZE (Section 3.1 and Fig. 3).

W2: Stage-removal ablations

Following the reviewer’s suggestion, we added stage-removal ablations (Tables 23 and 24). We compared F1 across 11 single-stage removals, seven view combinations, and a simple IoU-linking baseline with the same detection settings as WINZE. Removing whole-mask grouping or track filtering substantially reduced F1, and simple IoU linking scored below WINZE (Table 23). These results show that the improvement of WINZE comes from the newly added mechanism (Section 5.3).

W3a: Three volumes with sparse annotations

As the reviewer notes, the number of evaluation volumes is limited. This problem is due to the fact that natural-color tomography is a recent technique, thereby making annotated data scarce. To minimize this limitation, we prepared as many annotations as we could and evaluated annotated planes and densely annotated 3D ROIs in three cubes (Section 2.2). We also compared the methods on 11 publicly available volumes with ground truth (Table 3).

Regarding generalization, the design of WINZE suggests that it is applicable to a broad range of volume data that SAM 2 can segment. In the revised manuscript, we discuss this possibility and identify increasing the number of evaluation volumes as a future task for this domain (Section 6).

W3b: Low instance F1 and orthogonal-view precision

We agree that the absolute scores remain low. However, as the comparison with existing methods shows, this is a challenge for instance segmentation in natural-color tomography in general, not only for WINZE. WINZE achieved the highest plane-level F1 among the methods (Table 2) and the highest recall on the scored orthogonal planes (Table 8). These results indicate that WINZE offers an effective approach to improve instance segmentation in natural-color tomography. We state this conclusion in the revised Discussion (Section 6).

Q1: Fig. 3 and implementation details such as file names

Following the reviewer’s suggestion, we revised Fig. 3 and moved too specific details such as file names to Appendix C.1. The figure in the main text now shows the flow of information (Fig. 3).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely appreciate the reviewers’ valuable and insightful comments on our manuscript. We are glad that the reviewers found instance segmentation of natural-color tomography an underexplored and practically relevant problem. The concerns of the reviewers lie mainly in the following three points:

(i) the differences of WINZE from u-Segment3D and its methodological advances;

(ii) the contribution of each stage of WINZE (ablation study);

(iii) the scale and the validity of the evaluation.

Following the reviewers’ comments, we conducted additional experiments and analyses and addressed all of these points in the revised manuscript. Details are given in our response to each reviewer. The revision has significantly improved the manuscript and clarified the novelty and methodological advances of WINZE. The changes in the revised manuscript are highlighted in blue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.