- 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
- 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.
- 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
- 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.
- 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.
- 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
- 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?
- 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?
- 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.
- 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?
- 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
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.
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±0.35 RD mIoU and 45.30±0.36 RA mIoU. The corresponding structural metrics also show small run-to-run variability, with B0Err of 1.73±0.05 (RD) and 2.23±0.06 (RA), fragmentation of 0.222±0.009 and 0.272±0.011, and spurious-component counts of 1.20±0.04 and 1.49±0.05. The improvement is also preserved against the combined non-topological conditioning + CC-supervision control, which obtains 63.65±0.26 RD and 44.34±0.29 RA mIoU. Thus, TopoRadar retains gains of +1.27 and +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 LSC 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=5 retained components and runs 30 Sinkhorn iterations on transport matrices of size at most 5×5. Both LSC 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.