Pith. sign in

REVIEW 4 major objections 5 minor 31 references

FusionSort: Enhanced Cluttered Waste Segmentation with Advanced Decoding and Comprehensive Modality Optimization

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read FusionSort lifts cluttered waste segmentation to 73.12% mIoU, topping prior models on RGB, HSI, and fused inputs.

desk verdict The architecture is a plausible incremental decoder tweak, but the paper's own tables contradict its central fusion claim and its flagship MSWaste number, so the results as reported don't stand up. read the letter →

arxiv 2508.19798 v1 pith:O53CVQNZ submitted 2025-08-27 cs.CV

Muhammad Ali , Omar Ali AlSuwaidi This is my paper
classification cs.CV
keywords wastesegmentationhyperspectralimagingmultispectralRGB-HSIfusionMambaattentionPCAdimensionalityreductionsemanticclutteredscenes
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

FusionSort is a neural architecture for segmenting waste objects in cluttered, real-world streams, built on a SegFormer-style encoder with a redesigned decoder. The paper claims that adding a Comprehensive Attention Block, a Mamba attention branch, weighted feature fusion, and PCA-based RGB-HSI fusion improves segmentation accuracy across every tested modality. On SpectralWaste it reports mIoU of 67.26% with RGB alone, 61.02% with hyperspectral input, and 61.30% with fused RGB+HSI, beating lightweight and moderate baselines. On the nine-band MSWaste dataset it reports 73.12% IoU with only 6.7 million parameters, outperforming heavier encoder-decoder combinations. If these results hold under a controlled comparison, FusionSort would be a practical step toward automated sorting systems that can use whatever imaging hardware is available.

What carries the argument

The load-bearing mechanism is the decoder-side Comprehensive Attention Block: it splits the feature map into horizontal and vertical pooled branches, applies convolutional attention with sigmoid weighting, and runs a parallel Mamba state-space attention branch, then merges both via weighted attention fusion. Alongside it, the Data Fusion Block applies PCA to hyperspectral input to keep three channels of maximal variance, resizes them to match RGB, and concatenates the two streams. These components, with a combined Dice and cross-entropy loss, carry the claimed accuracy gains.

What would settle it

Retrain MiniNet-v2, SegFormer-B0, InternImage-T, and CMX-B0 on the same SpectralWaste train/validation/test split at 512x512 resolution for the same 60k iterations as FusionSort and compare mIoU; if the margins shrink to near zero or reverse, the superiority claim collapses. A second check would replace the PCA-reduced HSI input with randomly selected three bands, or with the full 224-band HSI, to see whether the PCA step itself is responsible for the fusion gain.

Watch

Extended reading notes

Core claim

The paper's central claim is that its FusionSort architecture, which keeps a transformer encoder and replaces the decoder with a Comprehensive Attention Block plus a parallel Mamba attention branch and weighted fusion, sets a new state of the art for waste object segmentation across RGB, hyperspectral, multispectral, and fused RGB+HSI inputs. The Data Fusion Block reduces 224-band hyperspectral data to three principal components before concatenating it with RGB, so the network sees both spectral signatures and spatial boundaries. On SpectralWaste, FusionSort reports 67.26% mIoU for RGB, 61.02% for PCA-reduced HSI, and 61.30% for the fused configuration, above MiniNet-v2, SegFormer-B0, Intern

Load-bearing premise

The comparison assumes every baseline in Table I was trained and evaluated under exactly the same protocol as FusionSort; if those numbers were taken from other papers instead, the claimed margins are not a controlled comparison.

Editorial extensions

If this is right

  • If the RGB-only result of 67.26% mIoU is reliable, sorting lines could get most of the benefit with ordinary RGB cameras and no hyperspectral hardware.
  • The fused RGB+HSI configuration's strong scores on Cardboard and Bag suggest spectral fusion helps precisely on materials that RGB tends to confuse.
  • At 73.12% IoU with 6.7 million parameters on MSWaste, the architecture would offer a lightweight alternative to ResNet- or Swin-based segmentation stacks of comparable accuracy.
  • The ablation pattern, with Mamba attention, Comprehensive Attention, and weighted fusion each contributing, implies the decoder modules are complementary rather than redundant.
  • A single architecture covering RGB, HSI, multispectral, and fused inputs means a waste-sorting deployment could adapt to whatever sensor suite is installed without changing the model family.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same PCA-based Data Fusion Block could in principle be applied to the nine-band MSWaste input or any future sensor with more than three channels, though the paper only demonstrates it on 224-band hyperspectral data.
  • On SpectralWaste the reported RGB-only mIoU (67.26%) is higher than the reported fused RGB+HSI mIoU (61.30%), so the paper does not itself show that fusion helps on that dataset; a controlled study isolating the fusion path would clarify whether the fusion block or the decoder is the main driver.
  • The heavy dependence on Mamba and coordinate attention suggests the decoder is learning directional and long-range context cues that matter most in cluttered scenes; testing the same modules on other cluttered segmentation tasks, such as medical or remote-sensing imagery, would show whether the gain generalizes beyond waste.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes FusionSort, an encoder-decoder architecture built on a SegFormer-style backbone, with three claimed contributions: a Comprehensive Attention Block combining coordinate-convolutional attention with a Mamba block, a Data Fusion Block that uses PCA to reduce hyperspectral inputs to three components before fusing them with RGB, and a combined Dice + cross-entropy loss (Eq. 1). The method is evaluated on the SpectralWaste dataset under RGB, HSI, and RGB+HSI configurations, and on the MSWaste multispectral dataset. The central claims are that the fused RGB+HSI configuration achieves the highest performance (61.30% mIoU) and that the approach outperforms existing methods by a significant margin on both datasets.

Significance. If the reported results were internally consistent, the paper would be a useful empirical contribution to waste sorting: it addresses a real application, covers RGB, HSI, multispectral, and fused inputs, and reports parameter counts as well as per-class IoU. The architecture description is detailed, and the paper includes a limitation paragraph acknowledging weaknesses on classes such as Tape. However, the central numerical claims are contradicted by the paper's own tables and by the arithmetic of the reported per-class scores. Because the claimed state-of-the-art result is the main contribution, these inconsistencies are load-bearing: the manuscript as submitted does not provide a reliable quantitative basis for its conclusions. The strengths are the breadth of the evaluation and the clear articulation of the proposed modules, not the evidence supporting the quantitative claims.

major comments (4)
  1. [IV-D.1, Table I] The sentence 'The fused RGB+HSI configuration of FusionSort achieves the highest overall performance with an mIoU of 61.30%' is directly contradicted by Table I, where the RGB-only FusionSort row reports 67.26% mIoU, 5.96 points higher. Since fusion is a central contribution and the abstract/title emphasize fusion, this is a load-bearing inconsistency. Additionally, in the HSI paragraph, CMX-B0 (56.6%) is discussed as an HSI result, but Table I places CMX-B0 in the RGB-Hyper3 block; thus the comparison is between FusionSort-HSI and a fused CMX-B0. Both statements need correction or the experiments need to be redone.
  2. [Table I] The per-class IoUs do not average to the reported mIoU for two rows. For FusionSort RGB+Hyper3: (66.09+72.48+86.12+24.68+55.79+56.61)/6 = 60.30, not 61.30. For FusionSort Hyper3: (61.94+73.27+83.73+34.18+64.13+53.54)/6 = 61.80, not 61.02. Only the RGB row averages correctly (67.26). Since mIoU is the paper's primary metric, the reported numbers for the HSI and fused configurations are unreliable.
  3. [IV-D.2, Tables II and III] The MSWaste result is inconsistent across the two tables. Table II reports MixVisionTrans.+FusionSort at 73.12% IoU, while Table III reports the complete 'All Modules' configuration at 70.51% mIoU on the same dataset. The text in IV-D.2 states 'an accuracy increase of up to 73%', while the ablation text states the highest mIoU is 70.51. A reader cannot determine the actual MSWaste mIoU or which number supports the claimed improvement.
  4. [IV-C, Table I] The baseline results (MiniNet-v2, SegFormer-B0, InternImage-T, CMX-B0) are presented without stating whether these models were retrained under the same protocol as FusionSort (same train/val/test split, 512x512 input, 60k iterations, same optimizer and loss). If the baseline numbers are taken from other publications rather than from a controlled in-house comparison, the 'significant margin' claim in the abstract is not supported. Please specify the training/evaluation protocol for every baseline or retrain them under identical conditions.
minor comments (5)
  1. [Throughout] There are numerous typos and inconsistent terms: 'FusioSort' vs 'FusionSort', 'Attetnion', 'ustilizing', 'traiined', 'mateerails', 'WasteMS', and the inconsistent use of 'Hyper3' / 'HSI' / 'RGB-H3'. Please standardize terminology.
  2. [References] References [10] and [11] list placeholder authors 'J. Doe and A. Roe', and reference [28] attributes the SpectralWaste dataset to 'M. Garcia and R. Torres' while reference [4] names the actual authors. This is not acceptable in a formal submission and must be corrected.
  3. [Eq. (1), III-C] The combined loss uses alpha=1 and beta=1 with no sensitivity analysis or justification. Since the loss trade-off can strongly affect segmentation results, please provide an ablation or a rationale.
  4. [Fig. 1] The heatmap in Fig. 1 reports aggregate values without clearly defining which 'Baseline' and 'Proposed' numbers are used. It should be reconciled with Tables I and II.
  5. [IV-A] The training iteration counts differ between datasets (60k for SpectralWaste, 28k for MSWaste) and the baseline training iterations are not reported. This makes it difficult to assess whether the comparisons are fair.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FusionSort reports empirical benchmark results; the only self-citation is contextual and not load-bearing.

full rationale

FusionSort is an empirical systems paper rather than a derived theoretical result. No step of the manuscript fits a parameter to a target quantity and then presents that quantity as a prediction. The PCA dimensionality reduction, the Comprehensive Attention Block, the Mamba attention, and the combined Dice/Cross-Entropy loss are architectural choices evaluated on held-out test sets (SpectralWaste and MSWaste); they are not constructed from the reported mIoU values. The hyperparameters α=β=1 are stated as fixed values, not fitted to maximize a reported metric. The sole self-citation, reference [2], is used in the Introduction to support the general premise that cluttered waste streams are challenging; it does not define any architecture, loss, or evaluation result, and the paper's central claim does not depend on it. There is no imported uniqueness theorem, no ansatz smuggled in via citation, and no renaming of a known empirical pattern as a new derivation. The paper does contain internal numerical inconsistencies (e.g., the RGB-only FusionSort mIoU of 67.26% in Table I exceeds the RGB+HSI mIoU of 61.30% that Section IV-D.1 calls 'highest overall performance', and some per-class averages do not match reported mIoUs). These are serious correctness/verification concerns, but they are not circularity: the claimed results do not reduce by construction to the paper's inputs. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central results depend on several hand-chosen hyperparameters (loss weights, PCA component count, Mamba width) that are not ablated or justified. The dataset annotation quality and the coherence of the training protocol are assumed. No new physical or conceptual entities are introduced.

free parameters (4)
  • alpha (Dice loss weight) = 1
    Set by hand in Eq. (1); no sensitivity analysis or ablation on the loss weighting.
  • beta (Cross Entropy loss weight) = 1
    Set by hand in Eq. (1); no sensitivity analysis.
  • PCA component count = 3
    The Data Fusion Block reduces HSI to 3 channels as a fixed choice; the paper does not ablate the number of components.
  • Mamba state expansion / local convolution width = not reported
    The Mamba Block uses state expansion and local convolution width parameters, but their values are not given, making replication impossible.
assumptions (4)
  • domain assumption SpectralWaste and MSWaste ground-truth annotations are correct and consistent.
    All metrics depend on label quality; the paper does not analyze annotation noise or label agreement.
  • ad hoc to paper PCA with 3 components retains the spectral information needed for waste-material discrimination.
    The component count is fixed without ablation or quantitative justification.
  • ad hoc to paper Combining Dice and cross-entropy loss with equal weights improves optimization.
    alpha=beta=1 is chosen by hand and not compared to other weightings.
  • domain assumption Mamba blocks can be inserted into a SegFormer decoder without destabilizing training.
    No training curves, convergence analysis, or sensitivity to Mamba width is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FusionSort: Enhanced Cluttered Waste Segmentation with Advanced Decoding and Comprehensive Modality Optimization." pith.science (2026). https://pith.science/paper/O53CVQNZ

@misc{pith2026250819798,
  author       = {Pith},
  title        = {Pith review of: FusionSort: Enhanced Cluttered Waste Segmentation with Advanced Decoding and Comprehensive Modality Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O53CVQNZ}},
  note         = {Machine review of arXiv:2508.19798}
}
read the original abstract

In the realm of waste management, automating the sorting process for non-biodegradable materials presents considerable challenges due to the complexity and variability of waste streams. To address these challenges, we introduce an enhanced neural architecture that builds upon an existing Encoder-Decoder structure to improve the accuracy and efficiency of waste sorting systems. Our model integrates several key innovations: a Comprehensive Attention Block within the decoder, which refines feature representations by combining convolutional and upsampling operations. In parallel, we utilize attention through the Mamba architecture, providing an additional performance boost. We also introduce a Data Fusion Block that fuses images with more than three channels. To achieve this, we apply PCA transformation to reduce the dimensionality while retaining the maximum variance and essential information across three dimensions, which are then used for further processing. We evaluated the model on RGB, hyperspectral, multispectral, and a combination of RGB and hyperspectral data. The results demonstrate that our approach outperforms existing methods by a significant margin.

Figures

Figures reproduced from arXiv: 2508.19798 by the authors.

Figure 1
Figure 1. Heatmap of segmentation performance (mIoU) for different data [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. FusionSort Architecture for Multimodal Waste Sorting. The architecture includes a hierarchical Backbone for feature extraction, and a Decoder with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison of segmentation results across three configurations of the proposed FusionSort architecture: FusionSort (RGB + HSI), FusionSort (RGB), [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: This Fig. presents a comprehensive analysis of the Fusionsort decoder’s feature maps. It represents feature maps obtained from comprehensive attention [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Results on the Multispectral Waste Dataset. The Fig. shows the Original Image (a), the Ground Truth segmentation map (b), and the Prediction by [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 30 canonical work pages

  1. [4]

    Spectralwaste dataset: Multimodal data for waste sorting automation,

    S. Casao, F. Pe ˜na, A. Sabater, R. Castill´on, D. Su´arez, E. Montijano, and A. C. Murillo, “Spectralwaste dataset: Multimodal data for waste sorting automation,” in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, code release for ”Spec- tralWaste Dataset: Multimodal Data for Waste Sorting Automation”

    work page 2024

  2. [28]

    The spectralwaste dataset: A benchmark for waste segmentation,

    M. Garcia and R. Torres, “The spectralwaste dataset: A benchmark for waste segmentation,” Waste Management AI, vol. 20, pp. 123–140, 2022

    work page 2022

  3. [1]

    Toward a cleaner and more sustainable world: A framework to develop and improve waste management through organizations, governments, and academia,

    R. Anuardo, M. Espuny, A. Costa, and O. Oliveira, “Toward a cleaner and more sustainable world: A framework to develop and improve waste management through organizations, governments, and academia,” Heliyon, vol. 8, no. 4, p. e09225, April 2022

    work page 2022

  4. [2]

    Enhanced segmentation of deformed waste objects in cluttered environments,

    M. Ali, O. Alsuwaidi, and S. Khan, “Enhanced segmentation of deformed waste objects in cluttered environments,” in Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods, ICPRAM 2024, Rome, Italy, February 24-26, 2024 . SCITEPRESS, 2024, pp. 570–581. [Online]. Available: https://doi.org/10.5220/0012424900003654

    work page doi:10.5220/0012424900003654 2024

  5. [3]

    Enhancing Environmental Monitoring through Multispectral Imaging: The WasteMS Dataset for Semantic Segmentation of Lakeside Waste

    Q. Zhu, N. Weng, L. Fan, and Y . Cai, “Enhancing environmental monitoring through multispectral imaging: The wastems dataset for semantic segmentation of lakeside waste,” 2024. [Online]. Available: https://arxiv.org/abs/2407.17028

    work page Pith review arXiv 2024

  6. [5]

    Maskgrasp: Mask-based grasping for scenes with multiple general real-world objects,

    J. Lee, J. Hur, I. Hwang, and Y . M. Kim, “Maskgrasp: Mask-based grasping for scenes with multiple general real-world objects,” in Pro- ceedings of the International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 3137–3144

    work page 2022

  7. [6]

    Recent advances of hyperspectral imaging technology and applications in agriculture,

    B. Lu, P. D. Dao, J. Liu, Y . He, and J. Shang, “Recent advances of hyperspectral imaging technology and applications in agriculture,” Remote Sensing, vol. 12, no. 16, 2020. [Online]. Available: https://www.mdpi.com/2072-4292/12/16/2659

    work page 2020

  8. [7]

    Trashnet: A dataset for garbage classification,

    G. Yang and G. Thung, “Trashnet: A dataset for garbage classification,” Stanford CS229 Project Report , 2016

    work page 2016

Show all 31 references
  1. [8]

    Taco: Trash annotations in con- text for litter classification,

    P. Proenc ¸a, A. Sim˜oes, and J. a. Araujo, “Taco: Trash annotations in con- text for litter classification,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , 2020

    2020

  2. [9]

    Floating waste: An autonomous boat dataset for water surface cleaning applications,

    X. Zhou and J. Smith, “Floating waste: An autonomous boat dataset for water surface cleaning applications,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) , 2021

    2021

  3. [10]

    Zerowaste: A large-scale dataset and benchmark for industrial waste sorting,

    J. Doe and A. Roe, “Zerowaste: A large-scale dataset and benchmark for industrial waste sorting,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , 2021

    2021

  4. [11]

    Zerowaste-v2: Enhanced industrial waste segmentation for visda 2022 challenge,

    ——, “Zerowaste-v2: Enhanced industrial waste segmentation for visda 2022 challenge,” in 2022 IEEE Winter Conference on Applications of Computer Vision (WACV) Workshops, 2022

    2022

  5. [12]

    Forest canopy chemistry with high spectral resolution remote sensing,

    F. Zagolski, V . Pinel, J. Romier, D. Alcayde, J. Fontanari, J. Gastellu- Etchegorry, G. Giordano, G. Marty, E. Mougin, and R. Joffre, “Forest canopy chemistry with high spectral resolution remote sensing,” Inter- national Journal of Remote Sensing , vol. 17, no. 6, pp. 1107–1...

    1996

  6. [13]

    Biophysical and biochemical sources of variability in canopy reflectance,

    G. P. Asner, “Biophysical and biochemical sources of variability in canopy reflectance,” Remote sensing of Environment , vol. 64, no. 3, pp. 234–253, 1998

    1998

  7. [14]

    Global food demand and the sustainable intensification of agriculture,

    D. Tilman, C. Balzer, J. Hill, and B. L. Befort, “Global food demand and the sustainable intensification of agriculture,” Proceedings of the national academy of sciences, vol. 108, no. 50, pp. 20 260–20 264, 2011

    2011

  8. [15]

    Ecological intensification of cereal production systems: yield potential, soil quality, and precision agriculture,

    K. G. Cassman, “Ecological intensification of cereal production systems: yield potential, soil quality, and precision agriculture,” Proceedings of the National Academy of Sciences, vol. 96, no. 11, pp. 5952–5959, 1999

    1999

  9. [16]

    Convolutional neural network for medical hyperspectral image classification with kernel fusion,

    Q. Huang, W. Li, and X. Xie, “Convolutional neural network for medical hyperspectral image classification with kernel fusion,” in BIBE 2018; International Conference on Biological Information and Biomedical Engineering. VDE, 2018, pp. 1–4

    2018

  10. [17]

    Change in the microenvironment of breast cancer studied by ftir imaging,

    S. Kumar, C. Desmedt, D. Larsimont, C. Sotiriou, and E. Goormaghtigh, “Change in the microenvironment of breast cancer studied by ftir imaging,” Analyst, vol. 138, no. 14, pp. 4058–4065, 2013

    2013

  11. [18]

    Water stress estimation of tomato canopy based on machine vision,

    K. Kurata and J. Yan, “Water stress estimation of tomato canopy based on machine vision,” in International Symposium on Plant Production in Closed Ecosystems 440 , 1996, pp. 389–394

    1996

  12. [19]

    Hyperspectral imaging based techniques in fluff sorting,

    G. Bonifazi and S. Serranti, “Hyperspectral imaging based techniques in fluff sorting,” in The 21st International Conference on Solid Waste Technology and Management. ICSWM, March 26–29 2006, pp. 740– 747

    2006

  13. [20]

    Hyperspectral imaging: A review on recent advances and applications,

    G. Lu and B. Fei, “Hyperspectral imaging: A review on recent advances and applications,” Journal of Biomedical Optics , vol. 19, no. 1, p. 010901, 2014

    2014

  14. [21]

    Fusion of hyperspectral imaging (hsi) and rgb for identification of soy- bean kernel damages using shufflenet with convolutional optimization and cross stage partial architecture,

    L. Zheng, M. Zhao, J. Zhu, L. Huang, J. Zhao, D. Liang, and D. Zhang, “Fusion of hyperspectral imaging (hsi) and rgb for identification of soy- bean kernel damages using shufflenet with convolutional optimization and cross stage partial architecture,” Frontiers in Plant Scienc...

    2023

  15. [22]

    Hyperspectral imaging in environmental monitoring: A review of recent developments and technological advances in compact field deployable systems,

    M. B. Stuart, A. J. S. McGonigle, and J. R. Willmott, “Hyperspectral imaging in environmental monitoring: A review of recent developments and technological advances in compact field deployable systems,” Sensors, vol. 19, no. 14, 2019. [Online]. Available: https://www.mdpi.com/...

    2019

  16. [23]

    Pandey, P

    C. Pandey, P. Sethy, and S. Behera, Hyperspectral Imagery Applications for Precision Agriculture , 02 2022, pp. 1–38

    2022

  17. [24]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,” in Advances in Neural Information Processing Systems , M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, ...

    2021

  18. [25]

    Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations,

    C. H. Sudre, W. Li, T. Vercauteren, S. Ourselin, and M. Jorge Car- doso, “Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations,” in Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: Third Int...

    2017

  19. [26]

    Cross-entropy loss functions: Theoretical analysis and applications,

    A. Mao, M. Mohri, and Y . Zhong, “Cross-entropy loss functions: Theoretical analysis and applications,” arXiv preprint arXiv:2304.07288, 2023

    2023 arXiv

  20. [27]

    MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark,

    M. Contributors, “MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark,” https://github.com/open- mmlab/mmsegmentation, 2020

    2020

  21. [29]

    Mininet: An efficient semantic segmentation convnet for real-time robotic applications,

    I. Alonso, L. Riazuelo, and A. C. Murillo, “Mininet: An efficient semantic segmentation convnet for real-time robotic applications,” IEEE Transactions on Robotics , vol. 36, no. 4, pp. 1340–1347, 2020

    2020

  22. [30]

    Internimage: Exploring large-scale vision foundation models with deformable convolutions,

    W. Wang, J. Dai, Z. Chen, Z. Huang, Z. Li, X. Zhu, X. Hu, T. Lu, L. Lu, H. Li et al., “Internimage: Exploring large-scale vision foundation models with deformable convolutions,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 14...

    2023

  23. [31]

    Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers,

    J. Zhang, H. Liu, K. Yang, X. Hu, R. Liu, and R. Stiefelhagen, “Cmx: Cross-modal fusion for rgb-x semantic segmentation with transformers,” IEEE Transactions on Intelligent Transportation Systems (TITS), vol. 24, pp. 14 679–14 694, 2023

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.