acceptodds
Under review as a conference paper at ICLR 2027

TopoRadar: Learning Persistent Topological Structure for Multi-View Radar Semantic Segmentation

Abstract

Radar semantic segmentation requires recovering foreground structure from sparse and fragmented observations. Range–angle, range–Doppler, and angle–Doppler projections provide complementary spatial and motion information, yet their different projection geometries can separate or merge object responses. Pixel-wise supervision does not explicitly constrain the resulting connectivity errors. We formulate multi-view radar semantic segmentation as a topological structural recovery problem, seeking to recover annotation-consistent foreground structure while respecting differences between views. We introduce TopoRadar, which couples persistent feature conditioning with component-level structural supervision. Its DuoPerLay module computes sublevel and superlevel cubical persistence from learned multi-view features, characterising connectivity and holes across filtration thresholds to condition the segmentation representation. A structural consistency loss uses optimal transport to align persistent component summaries between predictions and annotations within each output view, providing component-level supervision for foreground structure. On CARRADA, TopoRadar achieves 64.87% range–Doppler mIoU, outperforming the published TransRadar and recent Hyper- Radar results by 0.97 and 1.07 percentage points, respectively; in the range–angle view, it exceeds HyperRadar while remaining below TransRadar and MARSS. Relative to a TransRadar backbone re-trained under our protocol, TopoRadar reduces connected-component error by approximately 30% and spurious-component counts by approximately 35% across both views. A single-view adaptation achieves 83.17% mIoU on RADIal, improving TransRadar by 2.07 percentage points.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Radar Semantic Segmentation, Topological Deep Learning, Persistent Homology, Multi-View Radar Perception, Unbalanced Optimal Transport

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_mMk1OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TopoRadar for multi-view radar semantic segmentation. The method uses persistent homology to capture structural information, such as connectivity and fragmentation. It introduces DuoPerLay for topology-based feature conditioning and a Structural Consistency Loss based on unbalanced optimal transport. The authors evaluate the method on CARRADA and RADIal, including an analysis of structural consistency and component-level errors.

Critical strengths
  1. The paper addresses an important limitation of pixel-level segmentation metrics. Two predictions can have similar IoU but different connectivity and fragmentation. The proposed method explicitly considers these structural differences during learning. Table 5 shows reductions in connected-component errors, fragmentation, and spurious components in both RD and RA views.
  2. The paper provides useful ablation studies for the proposed components, including comparisons with simpler alternatives of similar model size. These experiments help show that the improvements are not simply due to additional model parameters. The evaluation on RADIal also demonstrates that the proposed approach can be adapted to a different radar representation.
Critical weaknesses
  1. The improvements in semantic segmentation are not consistent across classes and radar views. In Table 1, TopoRadar achieves 45.26% RA mIoU, compared with 47.5% for TransRadar and 46.97% for MARSS. The difference is particularly large for cyclists, where TopoRadar achieves only 9.5% IoU compared with 21.5% for TransRadar. Table 6 in the appendix also shows that some structural metrics become worse for RA cyclists. The authors acknowledge that the Structural Consistency Loss operates on class-agnostic foreground regions. However, these results suggest that improving foreground connectivity does not always improve semantic segmentation.
  2. The experimental results are reported from a single training run with a fixed random seed, as described in Appendix C. Since some of the improvements over existing methods are relatively small, it is unclear how consistent these gains would be across multiple runs. In addition, the best checkpoint is selected using RD validation mIoU, although the method is evaluated on both RD and RA. It would be useful to understand whether this selection strategy affects RA performance.
  3. The computational evaluation is somewhat limited. Table 10 compares TopoRadar only with TransRadar, while Tables 1 and 2 include several more recent methods. A computational comparison with these methods would help assess the accuracy versus efficiency tradeoff. Also, the paper does not provide a detailed computational breakdown of the proposed components. Although the Structural Consistency Loss is used only during training, its additional training overhead is not quantified.
Questions for authors
  1. Table 1 shows that RA cyclist segmentation is considerably worse than TransRadar. The authors attribute this partly to the class-agnostic structural loss. Have the authors considered class-aware structural supervision? Could this help improve cyclist segmentation while preserving the benefits in structural consistency?
  2. Appendix C states that the best checkpoint is selected based on RD validation mIoU. Why is RA performance not considered during checkpoint selection? Have the authors evaluated checkpoint selection using the average RD and RA mIoU?
  3. Could the authors report the mean and standard deviation of the results across multiple training runs? This would help establish the consistency of the reported improvements.
  4. Table 10 provides computational results only for TransRadar and TopoRadar. Could the authors include an efficiency comparison with more recent methods, such as HyperRadar and MARSS, under the same hardware and inference settings? Also, could they report the computational overhead of DuoPerLay and the additional training time introduced by the Structural Consistency Loss?
  5. The results on RADIal in Table 2 show that TopoRadar can be adapted to a different radar dataset. However, this does not establish cross-dataset generalization without retraining. Have the authors evaluated how well the learned representations transfer to a different radar dataset or sensor configuration with limited or no additional training?
2 comments

Q1: We thank the reviewer for the suggestion. Class-aware structural supervision could help address the RA cyclist weakness, although we have not yet tested it. We will include a controlled evaluation in the final version to assess whether it improves cyclist segmentation while preserving the overall structural gains.

Our current structural loss targets foreground connectivity rather than semantic class separation, which is supervised by the multiclass segmentation loss. Improved structural consistency therefore need not produce uniform gains across classes: Table 6 shows improvements for pedestrians and cars, whereas RA cyclists remain challenging. This does not establish the structural loss as the cause of the cyclist performance gap. The proposed evaluation will add class-specific structural matching alongside the existing foreground-union term and report both per-class IoU and structural metrics.

Q2: We select checkpoints using RD validation mIoU consistently for TopoRadar, the re-trained TransRadar backbone and all ablations, keeping model selection identical across the controlled comparisons. We have not evaluated selection using average RD and RA validation mIoU.

RA remains fully supervised during training; the RD-based criterion only determines which checkpoint is retained. We agree that applying an averaged criterion consistently to both the backbone and TopoRadar would also provide a fair comparison and could affect the reported results. We will clarify the selection protocol in Appendix C and acknowledge that sensitivity to this choice has not yet been assessed.

Q3: We report the requested multi-run evaluation in Table, given in reviewer xhTe response, using three independent training runs with seeds 40, 41, and 42. TopoRadar achieves 64.92±0.3564.92 \pm 0.35 RD mIoU and 45.30±0.3645.30 \pm 0.36 RA mIoU. The corresponding structural metrics also show small run-to-run variability, with B0Err of 1.73±0.051.73 \pm 0.05 (RD) and 2.23±0.062.23 \pm 0.06 (RA), fragmentation of 0.222±0.0090.222 \pm 0.009 and 0.272±0.0110.272 \pm 0.011, and spurious-component counts of 1.20±0.041.20 \pm 0.04 and 1.49±0.051.49 \pm 0.05. The improvement is also preserved against the combined non-topological conditioning + CC-supervision control, which obtains 63.65±0.2663.65 \pm 0.26 RD and 44.34±0.2944.34 \pm 0.29 RA mIoU. Thus, TopoRadar retains gains of +1.27+1.27 and +0.96+0.96 points in RD and RA, respectively, across independent runs, supporting the consistency of the reported improvements rather than a favourable single-seed result.

Q4: Table 10 was designed to isolate the incremental inference cost of TopoRadar relative to its TransRadar backbone under identical implementation and hardware settings. We additionally examined HyperRadar, for which the reported evaluation uses closely matched hardware and inference settings: both methods use a single RTX 5090, five-frame CARRADA inputs, and batch-size-1 inference. HyperRadar reports 1132.13 GFLOPs, 10.55 FPS, and 3.18 GB peak memory, whereas TopoRadar requires 1146.32 GFLOPs, 9.31 FPS, and 3.34 GB. Thus, HyperRadar is slightly more computationally efficient, while TopoRadar achieves higher CARRADA mIoU in both RD (64.87 vs. 63.8) and RA (45.26 vs. 44.4). Relative to TransRadar, DuoPerLay adds only 7.75 GFLOPs (+0.68%) and 0.09 GB peak memory (+2.77%), with throughput changing from 9.60 to 9.31 FPS (-3.0%). MARSS reports a compact 9.3M-parameter model, but does not provide directly comparable GFLOPs, FPS, or peak-memory measurements, so we avoid mixing heterogeneous runtime protocols.

We assess the training overhead of the Structural Consistency Loss in TopoRadar on an NVIDIA RTX 5090, holding the model architecture, batch size, five-frame input, optimiser, and data pipeline fixed. Training takes 0.55 s per iteration without LSCL_{\mathrm{SC}} and 0.62 s per iteration with it, a 12.73% increase in per-iteration training time. For each combination of view (RA or RD) and filtration (sublevel or superlevel), the loss uses at most K=5K=5 retained components and runs 30 Sinkhorn iterations on transport matrices of size at most 5×55 \times 5. Both LSCL_{\mathrm{SC}} and its unbalanced optimal transport (UOT) solver are used only during training, so they add no inference overhead.

Q5: We have not evaluated cross-dataset or cross-sensor transfer with limited or no additional training. Table 2 evaluates whether our design remains effective when adapted and trained on a different radar representation and annotation protocol.

Specifically, RADIal uses a single-view configuration with cross-view consistency disabled, whereas CARRADA uses RA/RD/AD inputs. These results support the applicability of the topology-aware design across these settings, but do not establish transfer of the learned representations. We will revise the manuscript, including the "transferability" wording in Appendix C, to make this distinction explicit.

Reviewer_mMk1OpenReview ↗

Response to Authors and Final Assessment

Thank you for the detailed responses and additional experiments. The multi-seed results and computational analysis address most of my concerns. The RA cyclist performance and checkpoint selection strategy remain limitations, but I appreciate the authors' clarifications.

Overall, the rebuttal has strengthened the paper. I will maintain my rating of 3.

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

This work focuses on radar semantic segmentation using range azimuth and range doppler heatmaps. The authors formulate the task as a topological structural recovery problem and introduce topology techniques to help encode these heatmaps and design the training objectives. Specifically, they use sublevel and superlevel filtrations during feature encoding and use an unbalanced optimal transport loss to align predicted objects with GT annotations. The paper also provides sufficient experimental results to demonstrate the effectiveness of the proposed method.

Critical strengths

The paper's motivation is clear.

It is interesting to regard radar-based segmentation as a tomography task. Both range azimuth and range doppler maps are grid-based which are naturally suitable for tomography. And I understand that it's usually difficult to identify targets in such heatmaps since there is only one intensity channel can be used and different targets have different levels of intensities, even within one target, the intensity could also be different. In addition, targets in these heatmaps often appear fragmented or incomplete due to radar's low resolution, missed detections and noise may also interfere these targets. Hence, it's reasonable to focus more on these stutural information instead of just considering pixel-level information.

The proposed modules are also well motivated. Targets in radar heatmaps often exhibit holes or irregular shapes, hence, the proposed DuoPerLay uses both sublevel and superlevel filtrations to capture birth/death persistence of connected components and holes. This enables the encoder to capture structural information at different levels.

Critical weaknesses

None. I think there is no critical weakness driving my decision, but I still have some concerns raised below.

Questions for authors
  1. Is it necessary to consider RA and RD maps together? Even for the same target, the annotation in RA and RD maps shows different width along the common range dimension. Hence, what's the motivation for training these two representations at the same time?

  2. What is L_pix in Figure 2?

  3. I feel that RD maps are more suitable for applying topological techniques as targets in RD maps usually exhibit clearer structures and the annotations are more consistent with the observed targets. However, for RA maps, there is often a gap between the original RA map and the annotations. The annotations usually cover smaller regions with more precise range and azimuth locations, but targets in RA maps often spread over much wider in both azimuth and range dimensions. In this case, can DuoPerLay still provide useful structural information when using RA maps alone? I am concerned that the structural information obtained from sublevel and superlevel filtrations in RA maps may not be consistent with GT annotations. Could this inconsistency introduce any undesired performance or is there any design helping to mitigate this inconsistency?

5 comments

Q1: RA and RD are trained jointly because they provide complementary projections of the same scene. We train jointly because RA provides azimuth information and RD provides radial-velocity information for the same scene. Following MVRSS and TransRadar, we encode RA, RD, and AD into a shared representation and predict separate RA and RD masks.

Each prediction is supervised by the annotation for its own view. The Structural Consistency Loss compares predictions and annotations within each view (Sec. 3.1 and Eq. (8)), allowing the two annotations to have different range extents and connectivity. The inherited multi-view loss LMVL_{\mathrm{MV}} directly couples the two predictions. For each class, it takes the maximum predicted probability over azimuth in RA and over Doppler in RD at each range bin, then applies a Huber penalty to the resulting range profiles. This encourages agreement across range bins as a soft constraint. If one view predicts a wider range extent, the extra bins can contribute to this penalty, while each view's annotation still supervises its own predicted extent.

On RADIal, a single-view adaptation with the cross-view loss disabled achieves 83.17%83.17\% mIoU, 2.072.07 percentage points above TransRadar (Table 2). This demonstrates that the method can be applied in a single-view setting. On CARRADA, we keep the joint setting because it is the benchmark protocol used by all compared methods, which keeps the results in Table 1 directly comparable.

Q2: LpixL_{\mathrm{pix}} in Fig. 2 denotes the baseline segmentation objective for the RA and RD predictions. It consists of the first three terms of Eq. (10):

Lpix=α1LCA+α2LSD+α3LMV.L_{\mathrm{pix}} = \alpha_1 L_{\mathrm{CA}} + \alpha_2 L_{\mathrm{SD}} + \alpha_3 L_{\mathrm{MV}}.

Here, LCAL_{\mathrm{CA}}, LSDL_{\mathrm{SD}}, and LMVL_{\mathrm{MV}} are the class-agnostic object loss, soft Dice loss, and multi-view consistency loss, respectively, with weights (α1,α2,α3)=(1,10,5)(\alpha_1,\alpha_2,\alpha_3)=(1,10,5). We use LpixL_{\mathrm{pix}} in Fig.~2 only as a compact notation for these baseline segmentation losses. We will add this definition to the caption of Fig. 2.

Q3: In our experiments, the gap between the spread of RA returns and the RA annotations does not prevent structural gains in RA, and two design choices limit its effect. Relative to the re-trained TransRadar backbone, the full model reduces RA foreground-union component-count error from 3.183.18 to 2.222.22, fragmentation from 0.380.38 to 0.270.27, and spurious components from 2.272.27 to 1.481.48 (Table 5), and adding DuoPerLay alone raises RA mIoU from 43.10%43.10\% to 44.18%44.18\% (Table 3).

A radar return can extend beyond the annotated object, so two choices keep the raw RA spread from being imposed on the prediction. First, DuoPerLay computes persistence on learned features after the three views are fused (Appendix C), not on the raw RA map. Second, the RA prediction is supervised by the RA annotation, which specifies the target extent; within the Structural Consistency Loss, the superlevel branch uses unbalanced optimal transport to allow partial matches between predicted and annotated foreground components with different masses.

The gains vary by class. For RA cyclists, the smallest targets, component-count error increases from 0.920.92 to 0.970.97 (Table 6). Our experiments use joint RA/RD/AD inputs, so whether DuoPerLay helps with RA input alone remains to be tested. We will discuss both limitations in Sec. 4.

Thank you for the detailed clarification. I understand that joint training of RA/RD/AD on CARRADA is for a fair comparison among all baseline methods, however, it still does not address my concern about why this joint training is required. As Doppler and azimuth provide different information, why do they complement each other in the segmentation task? I would also appreciate a better understanding of the motivation behind this choice.

Thank you for following up. Although the views provide different pieces of information, they describe the same underlying radar scene. Fundamentally, each resolved radar return (e.g., a signal reflected from an object) has a range, an azimuth and a Doppler value, while each 2-D view shows only two of these coordinates [1-3].

Segmentation using a subset of these coordinates, say range and Doppler, can therefore be ambiguous: different objects may share similar range and Doppler values while having different bearings (azimuths). This is analogous to multi-view computer vision, where two objects may overlap in one camera image but become distinguishable from another viewpoint. The additional view provides information that the first projection hides.

Example: Consider two road users at the same range. RA may show two distinct bearings, and RD may show two distinct radial velocities, while the pairing between the bearings and velocities remains ambiguous. AD can help resolve that ambiguity because it retains bearing and radial velocity together. This matters for segmentation. For instance, when a moving road user overlaps with static clutter in RA, Doppler provides evidence of motion, and AD can help associate that evidence with a bearing. Conversely, when returns from different bearings overlap in RD, RA gives the RD head spatial context for labelling that region

Thus, combining the views and training their segmentation heads jointly allows the model to bring together information about the same objects distributed across the 2-D views, helping it interpret ambiguous evidence in each view. This complementarity motivates our joint design.

[1] Ouaknine et al. Multi-View Radar Semantic Segmentation. ICCV, 2021.

[2] Dalbah et al. TransRadar: Adaptive-Directional Transformer for Real-Time Multi-View Radar Semantic Segmentation. WACV, 2024.

[3] Zia et al. “Learning Structurally Consistent Representations for Multi-view Radar Semantic Segmentation”. European Conference on Computer Vision (ECCV), 2026, pp. 359–376.

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

This paper proposes TopoRadar, which treats multi-view radar semantic segmentation as a problem of recovering topological structure. DuoPerLay incorporates sublevel and superlevel persistent homology into feature representations. A Structural Consistency Loss uses unbalanced optimal transport (UOT) to align connected components in predictions and ground-truth masks. Experiments on CARRADA and RADIal show that the method is effective, particularly in reducing foreground fragmentation and spurious connected components.

Critical strengths
  • The design combines feature conditioning with component-level supervision without requiring RA and RD to have the same topology. This is a reasonable choice for radar observations, where responses can split or merge differently across projections. It addresses structural properties that pixel-level accuracy alone may not capture.
  • Tables 3 and 5 show improvements in both mIoU and independent structural metrics for both views over a backbone re-trained under the same conditions. Table 9 also includes comparisons with approximately capacity-matched feature conditioning and an existing PH loss. These comparisons are valuable because they test whether the gains go beyond simply adding another module.
Critical weaknesses
  • There is a concern about the definition of the sublevel branch of the structural loss. As written, finite-persistence components of binary ground truth (GT) appear to contain only background just before they merge. Would this make their support mass, computed from foreground values, zero? The paper needs to clarify what this loss aligns, including how essential components that persist to the end of the filtration are handled.
Questions for authors
  • For sublevel components of binary GT, would computing mass from foreground values give zero mass? Please provide a concrete example showing the support region, mass, and contribution to UOT for both finite-persistence components and essential components that persist to the end of the filtration.
  • Please report the mean and variance across multiple random seeds, along with structural metrics, for the controls in Table 9. My assessment would improve if the results confirmed that the proposed method still offers additional gains over a baseline that combines non-topological conditioning with standard connected-component supervision.
2 comments

Q1: Yes. Both finite and essential sublevel components have zero foreground mass in a binary ground-truth (GT) mask. In the sublevel filtration described in Appendix C, background pixels enter first. We measure both components before the foreground pixels enter, so summing their GT pixel values gives zero. The sublevel loss can still penalise unmatched predicted mass, while the superlevel loss matches predictions to GT components with positive foreground mass. Consider this binary mask.

Yfg=[000111000].Y_{\mathrm{fg}} = \begin{bmatrix} 0 & 0 & 0 \\\\ 1 & 1 & 1 \\\\ 0 & 0 & 0 \end{bmatrix}.

The top and bottom rows form separate background components. The middle row enters at level N+1N+1 and joins them. Take the top row as the survivor of the tie. The bottom component ends at the merge (finite), while the top component survives it (essential). Following Appendix C, we take the finite component's pixels just before the merge, at level NN. For the essential component, we assign terminal level N+1N+1 and also take its pixels one step earlier, at level NN. The two supports are therefore the bottom and top rows, respectively. Each contains three zeros and enters unbalanced optimal transport (UOT) with only the numerical stabiliser κ=10−8\kappa=10^{-8} as its reference mass. Either tie outcome gives the same two masses.

Contribution to UOT: Consider a soft prediction with constant foreground probabilities of 0.100.10, 0.800.80, and 0.200.20 in the top, middle, and bottom rows, respectively. The finite predicted support is the bottom row just before the merge, with normalised mass 0.60/90.60/9. The essential predicted support covers the full grid at level NN, with normalised mass 3.30/93.30/9. A reference calculation of Eq. (7) with the Appendix C settings (ε=0.05\varepsilon=0.05, τ=0.8\tau=0.8, 30 Sinkhorn iterations) transports 3.19×10−53.19\times10^{-5} mass to each GT component. This can exceed κ\kappa because UOT allows transported mass to differ from the reference mass. Each GT component contributes about 0.000180.00018 to the weighted GT-side penalty. The weighted prediction-side penalty is 0.346170.34617 out of a total objective of 0.346540.34654. The penalty for leaving predicted mass unmatched therefore accounts for nearly all of the sublevel objective in this example.

Effect on segmentation in the RD view: We keep both filtrations in DuoPerLay and vary only the branches used in the Structural Consistency Loss. Adding the sublevel term to the superlevel loss improves RD mIoU by 0.62 percentage points and reduces both component-count error and spurious components.

Structural loss on CARRADA (RD)mIoU (%) ↑Betti-0 error ↓Spurious components ↓
Sublevel only63.852.051.45
Superlevel only64.251.821.25
Both branches (full TopoRadar)64.871.711.19

Structural metrics are computed on the foreground union following Appendix D.1. We will clarify the roles of the two output-level loss branches in Sec. 3.4 and state explicitly in Appendix C when we measure the essential component. DuoPerLay computes its filtrations on learned feature maps.

Q2: TopoRadar improves mean mIoU over the requested combined control by 1.27 percentage points in RD and 0.96 in RA, with lower component-count error, fragmentation and spurious counts in both views.

We trained TopoRadar, all Table 9 controls and the new combined control with seeds 40, 41 and 42 under the same protocol. Entries are mean ± SD (variance = SD² before rounding). RD = range-Doppler, RA = range-angle. mIoU is in %. B0 is component-count error. Frag. and Spur. denote fragmentation and spurious components. Structural metrics use the foreground union. The combined control uses parameter-matched non-topological conditioning and connected-component (CC) supervision. PH = persistent homology, SC = structural consistency.

MethodRD mIoURD B0RD Frag.RD Spur.RA mIoURA B0RA Frag.RA Spur.
Re-trained TransRadar62.13±0.302.47±0.070.312±0.0141.85±0.0643.13±0.333.19±0.090.382±0.0162.28±0.07
Matched non-topol. cond.62.85±0.282.35±0.060.300±0.0131.73±0.0543.57±0.313.04±0.080.368±0.0152.13±0.06
DuoPerLay H063.27±0.272.21±0.060.285±0.0121.59±0.0543.94±0.302.87±0.080.350±0.0141.97±0.06
DuoPerLay H0+H163.51±0.262.12±0.060.274±0.0121.49±0.0544.21±0.292.75±0.070.336±0.0141.86±0.06
CC w/o persistence62.99±0.292.26±0.070.291±0.0131.63±0.0643.64±0.322.93±0.080.358±0.0152.03±0.07
Existing PH loss63.11±0.282.18±0.060.283±0.0121.56±0.0543.75±0.312.85±0.080.347±0.0141.94±0.06
Backbone + SC loss63.24±0.272.09±0.060.272±0.0121.47±0.0543.89±0.302.72±0.070.333±0.0131.83±0.06
Combined control63.65±0.261.95±0.060.257±0.0111.37±0.0544.34±0.292.53±0.070.314±0.0131.70±0.06
TopoRadar64.92±0.351.73±0.050.222±0.0091.20±0.0445.30±0.362.23±0.060.272±0.0111.49±0.05
AuthorsOpenReview ↗

Follow-up on our response: requested combined control (Q2) and sublevel loss (Q1)

Dear Reviewer xhTe,

As the discussion period approaches its close, we would appreciate your feedback on whether our response addresses your concerns. In particular:

  • Requested controls and robustness (Q2): We evaluated TopoRadar, all Table 9 controls and the requested combined baseline across three random seeds, reporting mean ± SD and structural metrics. Against the combined non-topological conditioning + connected-component supervision baseline, TopoRadar improves mean mIoU by 1.27 percentage points in RD and 0.96 in RA, alongside lower component-count error, fragmentation and spurious-component counts in both views. This is the comparison you identified as a basis for improving your assessment.

  • Sublevel-loss definition (Q1): We provided a concrete worked example detailing the supports, masses and UOT contributions of both finite and essential components. We confirmed the zero GT foreground mass, clarified how the sublevel term penalises unmatched predicted mass, and will make this explicit in Sec. 3.4 and Appendix C.

Could you please let us know whether any concerns remain? If our response resolves them, we would be grateful if you would consider updating your assessment. We remain available to clarify any outstanding points.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.