acceptodds
Withdrawn from ICLR 2027

SAGE-RSS: Stage-Aligned Global–Local Evidence Refinement for Multi-View Radar Semantic Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Radar semantic segmentation (RSS) supports robust all-weather perception, yet radar frequency maps exhibit anisotropic structures, weak spectral responses, and temporal dependencies that generic image operators do not address. High-precision RSS must recover sparse target evidence while integrating context across heterogeneous radar projections. We present SAGE-RSS, a stage-aligned global–local evidence refinement framework. A Progressive Support Encoder (PSE) gathers multi-scale support through parallel channel groups and an identity path; a Global–Local Fusion Refiner (GLFR) combines an efficient transform-space branch with residual spatial enhancement; and a Context-Conditioned Axis-aware Decoder (CCAD) performs compact global readout and parallel multi-kernel reconstruction. On CARRADA, SAGE-RSS obtains 63.91% RD mIoU, 46.74% RA mIoU, and 55.33% mean-view mIoU. Under the same profiling protocol, the model runs at 33.30 FPS. The parallel branches and compact context aggregation provide real-time inference and make SAGE-RSS suitable for latency-sensitive and edge-side radar perception applications.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Radar Semantic Segmentation

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_AbQPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents SAGE-RSS for semantic segmentation across multiple radar views. The method is motivated by the observation that sparse radar evidence can be lost in different ways during encoding, fusion across views, and decoding. It introduces three dedicated modules to capture target support at multiple scales, combine global context with local details, and reconstruct structures specific to each view. Experiments on CARRADA show competitive results in both Range Doppler and Range Angle views.

Critical strengths

The paper presents a unified design logic. Encoding, fusion, and decoding address support at different scales, global and local context, and reconstruction adapted to each view, respectively. This organization by processing stage is intuitive. The combination of PSE, GLFR, and CCAD forms a coherent and relatively complete system.

Critical weaknesses
  1. The technical presentation is confusing and contains many inconsistencies between the text, equations, and figures. The raw input is defined as Xv∈RB×1×T×Hv×Wv\mathcal X^v\in\mathbb R^{B\times1\times T\times H_v\times W_v}, while PSE suddenly assumes X∈RB×C×H×W\mathbf X\in\mathbb R^{B\times C\times H\times W} without explaining the dimensional transition. Equation 1 computes the third PSE branch from [O1,X2,X3][O_1,X_2,X_3], whereas Figure 3 appears to use [O2,X3][O_2,X_3]. For CCAD, Equation 8 obtains the memory tokens through a fixed slice, whereas Figure 3 depicts a Top-K ranking. These discrepancies leave the actual architecture unclear and make the method difficult to reproduce.

  2. The ablation study is not described sufficiently. Table 3 only indicates which of PSE, GLFR, and CCAD are enabled, but does not explain what structure replaces each removed module. It also lacks internal ablations for the PSE identity route, kernels, and SE gates, the GLFR global and local branches and frequency weights, and the CCAD context branch, anchor settings, and directional kernels. It is unclear which specific design choices produce the reported gains.

  3. The evaluation does not directly support the core sparse evidence hypothesis. The motivation emphasizes weak pedestrian and cyclist returns, variation in target support, and the risk of erasing local peaks. The paper reports only aggregate mIoU/mDice, without IoU reported separately for pedestrians, cyclists, and cars or analyses stratified by target scale, sparsity, or signal strength. The qualitative evaluation is also limited to a single car example.

Questions for authors
  1. Could the authors clarify the tensor flow around PSE and resolve the inconsistencies between the equations and figures for PSE and CCAD?

  2. What architecture is used as the baseline in Table 3, and what replaces each removed module? Please also provide ablations of the main internal design choices.

  3. Could the authors report results for each class and analyses by target scale, sparsity, or signal strength, together with more examples of small target classes?

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

This paper proposes SAGE-RSS for multi-view radar semantic segmentation. The method includes three modules for feature encoding, global-local fusion, and decoding. The goal is to preserve weak target responses while combining information from different radar views. The authors evaluate the method on CARRADA and CARRADA-RAC and compare it with several existing approaches.

Critical strengths
  1. The paper addresses the challenge of preserving weak target responses during radar feature extraction and fusion. The proposed architecture considers this problem at different stages of the network. Table 3 shows that the three proposed modules contribute to segmentation performance, with the full model achieving better results than the individual configurations.
  2. The experimental results are encouraging. In Table 1, SAGE-RSS achieves the highest reported mIoU and mDice in both RD and RA views. Table 2 shows similar improvements on CARRADA-RAC. The authors also provide an efficiency comparison in Table 4, showing that the model achieves real-time inference on an RTX 3090 GPU.
Critical weaknesses
  1. The performance improvement comes with a substantial increase in model size. In Table 1, SAGE-RSS achieves 55.33% mean-view mIoU compared with 53.20% for AdaPKC-Net2, but uses 14.5M parameters instead of 6.3M. Table 3 shows the contribution of the individual modules, but does not report their parameter counts or computational costs. Therefore, it is difficult to determine how much of the improvement comes from the proposed architectural design rather than increased model capacity. A comparison with a baseline of similar size would strengthen the evaluation.
  2. The evaluation is limited to CARRADA and CARRADA-RAC. As mentioned in Section 4.3, CARRADA-RAC provides revised annotations of the same radar sequences. Although Table 2 shows that the improvements are consistent with these revised annotations, it does not establish generalization to a different radar dataset or sensing environment. This limits the conclusions about the robustness of the proposed approach.
Questions for authors
  1. In Table 1, SAGE-RSS uses more than twice as many parameters as AdaPKC-Net2 for an improvement of about 2 percentage points in mean-view mIoU. Could the authors compare the proposed method with a baseline having a similar number of parameters? Also, could the authors report the parameter counts or computational costs of the configurations in Table 3?
  2. The authors evaluate the method on CARRADA and CARRADA-RAC, which share the same underlying radar sequences. Is there any evaluation on an independent radar dataset? If not, could the authors discuss how well the method is expected to generalize to different radar environments or sensor configurations?
  3. Table 4 shows that SAGE-RSS achieves considerably higher FPS than AdaPKC-Net despite having more parameters and MACs. The authors attribute this to GPU-friendly operations. Could they provide more details about the benchmarking setup and explain the large difference in inference speed?
Reviewer_h4nPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SAGE-RSS, which preserves sparse radar evidence at each processing stage. SAGE-RSS combines multi-scale encoding with PSE, global-local fusion with GLFR, and context-conditioned axis-aware decoding with CCAD to predict RD and RA masks. On the CARRADA dataset, it achieved RD and RA mIoU scores of 63.91% and 46.74%, respectively, with an inference speed of 33.30 FPS.

Critical strengths
  • Experiments on CARRADA showed that SAGE-RSS outperformed AdaPKC-Net2 by 1.81 percentage points on RD and 2.44 percentage points on RA. It also showed improvements with the revised annotations.
  • Despite the increased computational cost, the method achieves a good balance between accuracy and measured inference speed, which is valuable for practical use.
Critical weaknesses
  • With the frequency-independent weights in Eq. (4), a full-spectrum FFT/IFFT gives F−1(F(X)⊙Wc)=X⊙Wc\mathcal{F}^{-1}(\mathcal{F}(X)\odot W_c)=X\odot W_c. The weight shape also differs from that shown in Figure 3. It is therefore unclear how the core global mixing mechanism works in the implementation.
  • Table 3 only compares the effects of adding modules. It does not include ablations using capacity-matched existing blocks or swapping module placements. Since the proposed method has about 2.3 times as many parameters as the AdaPKC models, it is not possible to tell whether the improvements come specifically from “stage alignment” or from increased model capacity.
Questions for authors
  • Please specify the FFT/IFFT variants used in GLFR, the shape of the weights and whether they depend on frequency, and how complex values are converted to real values.
  • Are the CCAD anchors the first NmN_m tokens in Eq. (8), or are they selected by Top-K as shown in Figure 3? In the former case, please explain why the anchors are limited to specific spatial positions and show their effectiveness compared with anchors that summarize the entire feature map.
  • Please describe the baseline model in Table 3 and provide ablation results in which the proposed modules are replaced with standard blocks, with matched parameter counts and training conditions. Results from swapping module placements are also important for assessing whether stage alignment is necessary as a design principle.

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

Related papers

Discussion (0)

Sign in to comment.