REVIEW 4 major objections 4 minor 12 references
Technical Report for ICRA 2025 GOOSE 2D Semantic Segmentation Challenge: Leveraging Color Shift Correction, RoPE-Swin Backbone, and Quantile-based Label Denoising Strategy for Robust Outdoor Scene Understanding
T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read A MaskDINO-based segmentation pipeline that adds color-shift correction, rotary position embeddings, and quantile-based label denoising reaches 0.848 mean IoU on the GOOSE outdoor test set.
desk verdict A genuine challenge win on GOOSE with a plausible pipeline, but the denoising step is described two ways and relies on the model's own errors, so the robustness claim for that component is under-supported. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing pieces are three modules grafted onto MaskDINO. CSEC first estimates a color-shift offset through a convolution with learned spatial offsets, then a color modulation module maps the image back to balanced colors by fusing self-correlation matrices of the original, darkened, and brightened feature maps with learned weights. RoPE-ViT replaces the default backbone by applying per-dimension rotations to query and key vectors in self-attention, with frequencies $\omega_i = 10000^{-2i/d}$, so relative patch positions enter the attention inner product directly. The quantile denoiser uses the current model to compute each training sample's pixel-wise error against its ground-truth mask and removes everything above the 97.5th percentile before continuing training.
What would settle it
Re-annotate the training samples the quantile filter removes above the 97.5th percentile error threshold, treating a human labeler as the reference; if most removed samples have human labels matching the original ground truth rather than the model's predictions, the denoising gain is an artifact of dropping hard examples rather than removing noise.
Extended reading notes
Core claim
On the official GOOSE test set, the paper reports a mean IoU of 0.848 for the full system: MaskDINO with a RoPE-ViT backbone, CSEC color correction, and quantile-based sample filtering. On the validation set the ablations are 87.92 without both RoPE and CSEC, 88.18 without RoPE, 88.72 without CSEC, 88.89 with both, and 89.13 after adding denoising. The authors read these numbers as evidence that color correction handles illumination inconsistencies, rotary position embeddings improve spatial generalization across camera resolutions and viewpoints, and removing the top 2.5% of high-error training samples stabilizes learning under label noise.
Load-bearing premise
The denoising step assumes that pixels the model itself predicts most incorrectly are mislabeled or anomalous, not valid hard examples, so discarding the top 2.5% cannot skew the training distribution.
Editorial extensions
If this is right
- MaskDINO users can adopt CSEC as a preprocessing stage and the RoPE backbone as a drop-in replacement without changing the segmentation head.
- On the GOOSE validation split, removing either RoPE or CSEC lowers mIoU, so the two modules contribute independently rather than only in combination.
- The quantile filter's validation gain from 88.89 to 89.13 suggests that even a simple hard sample-removal rule can improve training stability when annotation noise is present.
- Because the reported test mIoU is a weighted average (67% MuCAR-3, 24% ALICE, 6% Spot v2, 3% Spot v1), the headline number is driven mainly by two of the four platforms.
Reading between the lines
- Editorial inference: the quantile denoiser relies on the model's own training-set predictions, so high-error samples may be hard-but-valid examples; a manual re-annotation audit of the removed samples would settle whether the gain is real or a distribution-shift artifact.
- Editorial inference: the gap between 89.13 validation mIoU and 84.8 test mIoU points to residual domain shift between splits; per-platform reporting would show where color correction and rotary embeddings help most.
- Editorial inference: a soft-weighting version of the denoiser that downweights high-error pixels instead of deleting them is a natural testable extension that could preserve borderline training signals.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This technical report describes the ACVLAB submission to the ICRA 2025 GOOSE 2D Semantic Segmentation Challenge. The proposed method builds on MaskDINO and adds a Color Shift Estimation-and-Correction (CSEC) module, a RoPE-based Vision Transformer backbone, and a quantile-based label-denoising step that filters training samples. The authors report a test-set mIoU of 84.8% on the official GOOSE test set and provide validation-set ablations of the individual components.
Significance. The paper documents a competitive pipeline for semantic segmentation across heterogeneous outdoor robotic platforms, and the reported test result, if reliable, is of practical interest to the field-robotics community. The system-level design is coherent and the problem setting is clearly motivated. However, the paper's central robustness claims rest on the denoising component, whose empirical support is limited to a single validation-set comparison with a 0.24 mIoU gain and no error bars, and whose description is internally inconsistent. The absence of code, full training details, and test-set ablations further limits the value of the report beyond the competition outcome. The strengths are the concrete challenge result, the use of the official evaluation protocol, and the clear presentation of the architectural components.
major comments (4)
- [Abstract and Section II.C] The abstract and introduction describe the quantile-based strategy as 'downweighting the top 2.5% of highest-error pixels,' but Section II.C states that 'we remove samples that fall above the 97.5th percentile of the error rate distribution.' These operations are not equivalent: pixel-level downweighting preserves all training samples while reducing the loss contribution of selected pixels, whereas sample-level removal discards entire images. The paper must clarify which procedure was actually implemented and justify the discrepancy, because the claimed mechanism and the reported ablation depend on this distinction.
- [Section II.C and Table II] The denoising step selects training samples based on the model's own per-image error rate. This is a self-referential selection procedure: the model determines which training examples are 'noisy,' so the filtered training distribution is not independent of the model. High-error samples may be hard but valid examples, such as rare classes or boundary cases, and their removal can bias the training distribution and inflate validation mIoU without improving true robustness. With an improvement of only 0.24 mIoU (88.89 to 89.13) in Table II, no error bars, no randomization baseline, and no test-set ablation, the evidence is too weak to establish the claimed robustness to label noise.
- [Section III] The central test-set result (mIoU 84.8%) is reported as a single number without confidence intervals, without a per-platform breakdown despite the evaluation weights (67% MuCAR-3, 24% ALICE, 6% Spot v2, 3% Spot v1), and without a comparison of the ablated configurations from Tables I and II on the test set. This makes it impossible to determine whether the denoising step contributes to the test result or whether the validation-set gain is statistically meaningful. The authors should provide test-set results for all ablated configurations, ideally with multiple runs or an estimate of variance.
- [Section II.B and Fig. 1] The title and Figure 1 refer to a 'RoPE-Swin' backbone, but Section II.B states that the default Swin-L backbone is replaced by 'RoPE-ViT,' a Vision Transformer variant. This terminology inconsistency should be corrected, and the authors should clarify the exact architecture used, including whether the rotary positional embeddings are applied within a Swin-style or ViT-style model.
minor comments (4)
- [Section III] The experimental details are incomplete: the paper reports only the batch size and GPU type, and omits the number of training epochs, learning rate schedule, optimizer, data augmentation, and the specific point in training at which the denoising filter is applied.
- [Section II.A, Eq. (1)] In Equation (1), the summation index i and the domain of the feature maps x_i, weights w_i, and offsets Δp_i are not fully defined; please specify the ranges and shapes.
- [Tables I and II] Tables I and II report validation mIoU values differing by 0.5 to 1.0 points, but no variance estimates or repeated runs are given; the authors should state whether these differences are stable across seeds.
- [References] Reference [9] for RoPE and reference [2] for MaskDINO would benefit from complete venue information (e.g., ECCV or CVPR pages) to aid readers.
Circularity Check
The quantile-based label denoising claim is self-confirming: the noise set is defined by the model's own prediction errors, so the reported validation gain does not independently establish robustness to label noise.
-
fitted input called prediction
[Section II.C (Quantile-based Label Denoising Strategy) and Table II in Section III]
"Specifically, we use the existing model to predict each training sample and calculate the pixel-wise error rate between the predicted mask and the corresponding ground truth mask. After statistically analyzing the error rates of all training samples, we remove samples that fall above the 97.5th percentile of the error rate distribution, retaining the relatively normal samples below the 97.5th percentile."
The noisy-sample set is selected by the model's own pixel-wise prediction errors rather than by any independent annotation-quality signal. The paper then trains on the complement and, in Table II, reports that this filtering 'effectively enhances training data quality and segmentation capability' (validation mIoU rises from 88.89% to 89.13%). Because the filter is a function of the model's prior competence, the improvement can be explained as the model avoiding training samples it already finds difficult, not as verified removal of label noise. The claimed robustness-to-label-noise conclusion is therefore self-confirming; no test-set ablation or independent noise verification is provided.
full rationale
The central test-set number (84.8% mIoU) is an externally measured benchmark result, and the CSEC and RoPE components are imported from prior work without a self-citation chain, so the pipeline as a whole is not circular. The circularity is localized to the quantile-based denoising ablation: Section II.C defines the noise set as samples whose pixel-wise error, computed by the paper's own 'existing model', exceeds the 97.5th percentile, and Table II then presents the improved validation mIoU as evidence that this 'effectively enhances training data quality.' Since the filter is generated by the model's own errors, the improvement is confounded with the model's prior difficulty on those samples and does not independently verify label-noise removal. The abstract's description (downweighting the top 2.5% of pixels) also does not match the method (removing whole samples above the 97.5th percentile), further weakening the denoising attribution. No test-set ablation or independent noise labels are provided, so the denoising contribution is only partially supported. This warrants a score of 6: partial circularity in one component while the overall benchmark result remains externally measured.
Assumptions & free parameters
free parameters (1)
- Denoising percentile threshold =
0.975 (97.5th percentile of sample error rates)
assumptions (3)
- domain assumption GOOSE ground-truth annotations are mostly correct and the remaining errors can be identified by the model's own prediction errors.
- domain assumption The CSEC color correction module improves segmentation across all four robot platforms without introducing new domain-specific artifacts.
- domain assumption RoPE-ViT can be swapped into MaskDINO and retains its pretrained representation strength.
Cite this review
Pith. "Pith review of Technical Report for ICRA 2025 GOOSE 2D Semantic Segmentation Challenge: Leveraging Color Shift Correction, RoPE-Swin Backbone, and Quantile-based Label Denoising Strategy for Robust Outdoor Scene Understanding." pith.science (2026). https://pith.science/paper/QZH5TYXG
@misc{pith2026250506991,
author = {Pith},
title = {Pith review of: Technical Report for ICRA 2025 GOOSE 2D Semantic Segmentation Challenge: Leveraging Color Shift Correction, RoPE-Swin Backbone, and Quantile-based Label Denoising Strategy for Robust Outdoor Scene Understanding},
year = {2026},
howpublished = {\url{https://pith.science/paper/QZH5TYXG}},
note = {Machine review of arXiv:2505.06991}
}
read the original abstract
This report presents our semantic segmentation framework developed by team ACVLAB for the ICRA 2025 GOOSE 2D Semantic Segmentation Challenge, which focuses on parsing outdoor scenes into nine semantic categories under real-world conditions. Our method integrates a Swin Transformer backbone enhanced with Rotary Position Embedding (RoPE) for improved spatial generalization, alongside a Color Shift Estimation-and-Correction module designed to compensate for illumination inconsistencies in natural environments. To further improve training stability, we adopt a quantile-based denoising strategy that downweights the top 2.5\% of highest-error pixels, treating them as noise and suppressing their influence during optimization. Evaluated on the official GOOSE test set, our approach achieved a mean Intersection over Union (mIoU) of 0.848, demonstrating the effectiveness of combining color correction, positional encoding, and error-aware denoising in robust semantic segmentation.
Figures
Reference graph
Works this paper leans on
-
[1]
The goose dataset for perception in unstructured environments,
P. Mortimer, R. Hagmanns, M. Granero, T. Luettel, J. Petereit, and H.-J. Wuensche, “The goose dataset for perception in unstructured environments,” 2023. [Online]. Available: https://arxiv.org/abs/2310. 16788
work page 2023
-
[2]
Mask dino: Towards a unified transformer-based framework for object detection and segmentation,
F. Li, H. Zhang, H. xu, S. Liu, L. Zhang, L. M. Ni, and H.-Y . Shum, “Mask dino: Towards a unified transformer-based framework for object detection and segmentation,” 2022. [Online]. Available: https://arxiv.org/abs/2206.02777
arXiv 2022
-
[3]
U-net: Convolutional networks for biomedical image segmentation,
O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” 2015. [Online]. Available: https://arxiv.org/abs/1505.04597
arXiv 2015
-
[4]
Encoder-decoder with atrous separable convolution for semantic image segmentation,
L.-C. Chen, Y . Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder-decoder with atrous separable convolution for semantic image segmentation,” 2018. [Online]. Available: https://arxiv.org/abs/ 1802.02611
arXiv 2018
-
[5]
Pyramid scene parsing network,
H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid scene parsing network,” 2017. [Online]. Available: https://arxiv.org/abs/1612.01105
arXiv 2017
-
[6]
Masked-attention mask transformer for universal image segmenta- tion,
B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar, “Masked-attention mask transformer for universal image segmenta- tion,” 2022
2022
-
[7]
Pem: Prototype-based efficient maskformer for image segmentation,
N. Cavagnero, G. Rosi, C. Cuttano, F. Pistilli, M. Ciccone, G. Averta, and F. Cermelli, “Pem: Prototype-based efficient maskformer for image segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 804–15 813
work page 2024
-
[8]
Color shift estimation-and- correction for image enhancement,
Y . Li, K. Xu, G. P. Hancke, and R. W. Lau, “Color shift estimation-and- correction for image enhancement,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024
work page 2024
Show all 12 references
-
[9]
Rotary position embedding for vision transformer,
B. Heo, S. Park, D. Han, and S. Yun, “Rotary position embedding for vision transformer,” inEuropean Conference on Computer Vision (ECCV), 2024
2024
-
[10]
Swin transformer: Hierarchical vision transformer using shifted windows,
Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021
2021
-
[11]
Attention is all you need,
A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,”
-
[2023]
Available: https://arxiv.org/abs/1706.03762 Winners of the GOOSE 2D Semantic Segmentation Challenge at the IEEE ICRA Workshop on Field Robotics 2025
[Online]. Available: https://arxiv.org/abs/1706.03762 Winners of the GOOSE 2D Semantic Segmentation Challenge at the IEEE ICRA Workshop on Field Robotics 2025
2025 arXiv
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.