Pith. sign in

REVIEW 2 major objections 3 minor 38 references

TNet: Terrace Convolutional Decoder Network for Remote Sensing Image Semantic Segmentation

T0 review · 2 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read TNet claims that a decoder built from only convolutions and additions can match Transformer- and Mamba-augmented decoders on three remote sensing segmentation benchmarks.

desk verdict A simple conv-only decoder with competitive reported numbers; the idea is plausible but the abstract alone does not support the central claim. read the letter →

arxiv 2508.04061 v2 pith:ASTAOIO5 submitted 2025-08-06 cs.CV

Chengqian Dai , Yonghong Guo , Hongzhao Xiang , Yigui Luo This is my paper
classification cs.CV
keywords TNetsemanticsegmentationremotesensingconvolutionaldecoderprogressivefeaturefusionmulti-resolutioncontextVaihingenLoveDA
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

TNet asks whether a segmentation decoder can drop Transformer and Mamba modules and still capture global context. Its answer is yes: by progressively adding low-resolution features into higher-resolution stages, the decoder learns convolution kernels that blend global and local information. Using a ResNet-18 encoder, TNet reaches mean intersection-over-union values of 85.35% on Vaihingen, 87.05% on Potsdam, and 52.19% on LoveDA. The practical point is that cross-resolution context, usually bought with attention, can instead be encoded by simple stage-wise convolution and addition.

What carries the argument

Terrace Convolutional Decoder: a decoder whose stages are arranged like terraces, with each stage taking a coarser, globally informative feature map and adding it into a finer, locally detailed feature map before convolution. The operation is a learned convolution plus elementwise addition, and it is the mechanism the paper uses to learn spatially-aware kernels that fuse global and local context without attention.

What would settle it

Train TNet-R and a plain ResNet-18 UNet from identical seeds on the Vaihingen training split and evaluate both on the official test images; if the terrace decoder does not beat the plain decoder by more than seed-to-seed variation, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that global context in semantic segmentation does not require attention modules inside the decoder. TNet progressively integrates low-resolution features, which carry global context, into higher-resolution features, which carry local detail, across decoding stages; each stage learns spatially-aware convolution kernels that combine multiple scales through convolution followed by addition. Compared with UNet-style decoders that bolt on Transformer or Mamba modules, TNet keeps the decoder to convolution and addition only. On the Vaihingen, Potsdam, and LoveDA remote sensing benchmarks, the ResNet-18 variant reaches 85.35%, 87.05%, and 52.19% mean IoU, respectively,

Load-bearing premise

The reported mIoU values only support the competitiveness claim if every comparison was run on the same official train/test partitions with the same evaluation rules.

Editorial extensions

If this is right

  • A decoder with only standard convolution and addition can reach competitive accuracy on aerial-image segmentation, so attention modules are not required in the decoder for these benchmarks.
  • The reported numbers on Vaihingen, Potsdam, and LoveDA show the design is not limited to one dataset and transfers across different remote sensing scenes.
  • Because the decoder uses only convolution and addition, it has a smaller operation set than attention-based decoders, which matters when processing large aerial images.
  • The design can be paired with a standard ResNet-18 encoder and trained end to end with a normal segmentation pipeline.

Reading between the lines

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

  • The terrace fusion is described independently of the encoder, so a direct next test is to swap ResNet-18 for a stronger or weaker backbone and see whether the accuracy advantage persists.
  • An ablation that replaces the terrace connections with ordinary lateral skip connections would isolate how much of the gain comes from cross-resolution addition rather than from the added parameter count.
  • Since all decoder operations are standard, porting the decoder to latency-sensitive or on-board settings is plausible; the paper reports efficiency but leaves edge-device latency as an open empirical question.
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

2 major / 3 minor

Summary. The paper proposes TNet (Terrace Convolutional Decoder Network), a semantic segmentation architecture that uses only convolution and addition operations in the decoder to progressively fuse low-resolution global context with higher-resolution local details across decoding stages. The authors implement TNet with a ResNet-18 encoder (TNet-R) and report mean Intersection-over-Union (mIoU) results on three remote sensing benchmarks: 85.35% on ISPRS Vaihingen, 87.05% on ISPRS Potsdam, and 52.19% on LoveDA, claiming competitive performance and high computational efficiency. Code is stated to be publicly available.

Significance. If the reported results are reproducible and the comparisons are fair, this work would provide a simple and efficient alternative to heavier attention- or Mamba-based decoder designs, suggesting that progressive multi-resolution fusion with basic operations can match more complex mechanisms. The architectural idea is plausible and of potential practical value for remote sensing segmentation. However, the significance can only be assessed if the full experimental protocol and controlled baselines are provided; the abstract alone does not establish the claimed competitiveness.

major comments (2)
  1. [Abstract] The central claim of competitive performance rests on three mIoU numbers, but the abstract provides no evaluation protocol: train/validation splits, definition of mIoU (e.g., whether the background class is included or ignored), number of runs, data augmentation, test-time augmentation, or class weighting. Without these details, the numbers are not directly comparable to prior published results, and the claim is not falsifiable from the abstract. If the full paper does not specify these in its experimental section, the results must be considered unverified. This is load-bearing because every subsequent claim of competitiveness depends on the comparability of these numbers.
  2. [Abstract] The claim of 'competitive performance' is not supported by any baseline comparison in the abstract. The paper should include a controlled ablation (or at least a comparison) using the same ResNet-18 encoder with a standard UNet decoder and a representative attention-based decoder under identical training settings. Without such a controlled comparison, the reported mIoU could be driven by training hyperparameters, data augmentation, or optimization details rather than by the proposed terrace decoder. This concern also applies to the claim of computational efficiency, which is stated without reporting parameters, FLOPs, or latency.
minor comments (3)
  1. [Abstract] The abstract states 'high computational efficiency' but does not report any quantitative measure (parameters, FLOPs, inference time). Please provide these numbers or a reference to a table in the full paper.
  2. [Abstract] The term 'terrace' is not defined. The abstract should briefly explain what makes the decoder 'terrace-like' (e.g., stepwise integration of feature maps at descending scales) to help readers understand the core idea.
  3. [Abstract] The abstract does not mention the input resolution, number of training epochs, or hardware; these details are needed to interpret the efficiency claim and to enable reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical benchmark evaluation is self-contained.

full rationale

The paper is an empirical architecture proposal. The central claim is that a specific decoder (TNet) achieves competitive mIoU on three standard benchmarks. This claim is supported by direct experimental evaluation, not by a derivation chain that reduces to its own inputs. There are no equations in the available text that define one quantity in terms of another, no parameter fitted to a subset and then renamed a prediction, and no self-citation invoked as load-bearing evidence. The architecture is presented as a novel design, and its performance is measured against external benchmarks, making the result externally falsifiable. The abstract's omission of experimental protocol details (splits, number of runs, augmentation) is a reproducibility concern and a potential threat to comparability, but it is not circular reasoning under the defined taxonomy. No step in the paper's argument reduces to its own assumptions or to a self-citation chain. Therefore, the circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim depends on the evaluation protocol and the choice of architecture hyperparameters, none of which are specified in the abstract.

free parameters (2)
  • Decoder stage channel widths and kernel sizes
    Not reported in the abstract; these hyperparameters are typically chosen by hand and tuned on validation sets.
  • Training hyperparameters (learning rate, batch size, epochs, etc.)
    Not mentioned in the abstract; required to reproduce the reported mIoU values.
assumptions (3)
  • domain assumption The reported mIoU scores use the standard train/validation/test splits of the three benchmarks.
    Without this, the numbers are not comparable to prior work. The abstract does not state the protocol.
  • domain assumption mIoU is an adequate proxy for semantic segmentation quality in remote sensing.
    The paper uses mIoU as its primary metric, which is standard but does not capture all aspects of segmentation quality.
  • domain assumption The three datasets cover enough variety to generalize the architecture's effectiveness.
    The abstract evaluates on Vaihingen, Potsdam, and LoveDA, but without full analysis it is unknown whether results extend to other remote sensing domains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TNet: Terrace Convolutional Decoder Network for Remote Sensing Image Semantic Segmentation." pith.science (2026). https://pith.science/paper/ASTAOIO5

@misc{pith2026250804061,
  author       = {Pith},
  title        = {Pith review of: TNet: Terrace Convolutional Decoder Network for Remote Sensing Image Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ASTAOIO5}},
  note         = {Machine review of arXiv:2508.04061}
}
read the original abstract

In remote sensing, most segmentation networks adopt the UNet architecture, often incorporating modules such as Transformers or Mamba to enhance global-local feature interactions within decoder stages. However, these enhancements typically focus on intra-scale relationships and neglect the global contextual dependencies across multiple resolutions. To address this limitation, we introduce the Terrace Convolutional Decoder Network (TNet), a simple yet effective architecture that leverages only convolution and addition operations to progressively integrate low-resolution features (rich in global context) into higher-resolution features (rich in local details) across decoding stages. This progressive fusion enables the model to learn spatially-aware convolutional kernels that naturally blend global and local information in a stage-wise manner. We implement TNet with a ResNet-18 encoder (TNet-R) and evaluate it on three benchmark datasets. TNet-R achieves competitive performance with a mean Intersection-over-Union (mIoU) of 85.35\% on ISPRS Vaihingen, 87.05\% on ISPRS Potsdam, and 52.19\% on LoveDA, while maintaining high computational efficiency. Code is publicly available.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 24 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Chen, Y.; Fang, P.; Zhong, X.; Yu, J.; Zhang, X.; and Li, T. 2024. Hi-ResNet: Edge Detail Enhancement for High-Resolution Remote Sensing Segmentation. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17: 15024--15040

    work page 2024

  4. [4]

    G.; Kirillov, A.; and Girdhar, R

    Cheng, B.; Misra, I.; Schwing, A. G.; Kirillov, A.; and Girdhar, R. 2022. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1290--1299

    2022

  5. [5]

    Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-Fei, L. 2009. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, 248--255

    work page 2009

  6. [6]

    Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929

    arXiv 2020

  7. [7]

    Fu, J.; Liu, J.; Tian, H.; Li, Y.; Bao, Y.; Fang, Z.; and Lu, H. 2019. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 3146--3154

    work page 2019

  8. [8]

    Fu, Y.; Li, W.; Fan, S.; Jiang, Y.; and Bai, H. 2023. CAL-Net: Conditional attention lightweight network for in-orbit landslide detection. IEEE Transactions on Geoscience and Remote Sensing, 61: 1--15

    work page 2023

Show all 38 references
  1. [9]

    Gao, F.; Fu, M.; Cao, J.; Dong, J.; and Du, Q. 2025. Adaptive Frequency Enhancement Network for Remote Sensing Image Semantic Segmentation. IEEE Transactions on Geoscience and Remote Sensing, 63: 1--15

    2025

  2. [10]

    Gu, A.; and Dao, T. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752

    2023 arXiv

  3. [11]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 770--778

    2016

  4. [12]

    Himeur, Y.; Rimal, B.; Tiwary, A.; and Amira, A. 2022. Using artificial intelligence and data fusion for environmental monitoring: A review and future perspectives. Information Fusion, 86: 44--75

    2022

  5. [13]

    Hwang, G.; Jeong, J.; and Lee, S. J. 2024. SFA-Net: Semantic feature adjustment network for remote sensing image segmentation. Remote Sensing, 16(17): 3278

    2024

  6. [14]

    Li, L.; Yi, J.; Fan, H.; and Lin, H. 2025. A Lightweight Semantic Segmentation Network Based on Self-Attention Mechanism and State Space Model for Efficient Urban Scene Segmentation. IEEE Transactions on Geoscience and Remote Sensing, 63: 1--15

    2025

  7. [15]

    Li, R.; Zheng, S.; Zhang, C.; Duan, C.; Wang, L.; and Atkinson, P. M. 2021. ABCNet: Attentive bilateral contextual network for efficient semantic segmentation of Fine-Resolution remotely sensed imagery. ISPRS journal of photogrammetry and remote sensing, 181: 84--98

    2021

  8. [16]

    Lin, T.-Y.; Doll \'a r, P.; Girshick, R.; He, K.; Hariharan, B.; and Belongie, S. 2017. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2117--2125

    2017

  9. [17]

    Liu, M.; Dan, J.; Lu, Z.; Yu, Y.; Li, Y.; and Li, X. 2024. CM-UNet: Hybrid CNN-Mamba UNet for remote sensing image semantic segmentation. arXiv preprint arXiv:2405.10530

    2024 arXiv

  10. [18]

    Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; and Guo, B. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, 10012--10022

    2021

  11. [19]

    Liu, Z.; Mao, H.; Wu, C.-Y.; Feichtenhofer, C.; Darrell, T.; and Xie, S. 2022. A ConvNet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2022

  12. [20]

    Loshchilov, I.; and Hutter, F. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101

    2017 arXiv

  13. [21]

    Ma, X.; Zhang, X.; and Pun, M.-O. 2024. RS3Mamba: Visual State Space Model for Remote Sensing Image Semantic Segmentation. IEEE Geoscience and Remote Sensing Letters, 21: 1--5

    2024

  14. [22]

    Ronneberger, O.; Fischer, P.; and Brox, T. 2015. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, 234--241. Springer

    2015

  15. [23]

    Sandler, M.; Howard, A.; Zhu, M.; Zhmoginov, A.; and Chen, L.-C. 2018. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 4510--4520

    2018

  16. [24]

    Strudel, R.; Garcia, R.; Laptev, I.; and Schmid, C. 2021. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, 7262--7272

    2021

  17. [25]

    Tan, M.; and Le, Q. 2019. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, 6105--6114. PMLR

    2019

  18. [26]

    Tan, M.; and Le, Q. 2021. Efficientnetv2: Smaller models and faster training. In International conference on machine learning, 10096--10106. PMLR

    2021

  19. [27]

    Vaihingen, I.; and Potsdam, I. 2024. ISPRS Vaihingen and ISPRS Potsdam datasets

    2024

  20. [28]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  21. [29]

    Wang, J.; Zheng, Z.; Ma, A.; Lu, X.; and Zhong, Y. 2021 a . Love DA : A Remote Sensing Land-Cover Dataset for Domain Adaptive Semantic Segmentation

    2021

  22. [30]

    Wang, L.; Li, D.; Dong, S.; Meng, X.; Zhang, X.; and Hong, D. 2024. PyramidMamba: Rethinking pyramid feature fusion with selective space state model for semantic segmentation of remote sensing imagery. arXiv preprint arXiv:2406.10828

    2024 arXiv

  23. [31]

    Wang, L.; Li, R.; Duan, C.; Zhang, C.; Meng, X.; and Fang, S. 2022 a . A Novel Transformer Based Semantic Segmentation Scheme for Fine-Resolution Remote Sensing Images. IEEE Geoscience and Remote Sensing Letters, 19: 1--5

    2022

  24. [32]

    Wang, L.; Li, R.; Wang, D.; Duan, C.; Wang, T.; and Meng, X. 2021 b . Transformer meets convolution: A bilateral awareness network for semantic segmentation of very fine resolution urban scene images. Remote Sensing, 13(16): 3065

    2021

  25. [33]

    Wang, L.; Li, R.; Zhang, C.; Fang, S.; Duan, C.; Meng, X.; and Atkinson, P. M. 2022 b . UNetFormer: A UNet-like transformer for efficient semantic segmentation of remote sensing urban scene imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 190: 196--214

    2022

  26. [34]

    Wightman, R. 2019. PyTorch Image Models. https://github.com/huggingface/pytorch-image-models

    2019

  27. [35]

    Wu, H.; Huang, P.; Zhang, M.; Tang, W.; and Yu, X. 2023. CMTFNet: CNN and Multiscale Transformer Fusion Network for Remote-Sensing Image Semantic Segmentation. IEEE Transactions on Geoscience and Remote Sensing, 61: 1--12

    2023

  28. [36]

    Zhang, X.; Liu, Y.; Lin, Y.; Liao, Q.; and Li, Y. 2024. Uv-sam: Adapting segment anything model for urban village identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 22520--22528

    2024

  29. [37]

    Zhu, L.; Liao, B.; Zhang, Q.; Wang, X.; Liu, W.; and Wang, X. 2024. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417

    2024 arXiv

  30. [38]

    Zou, X.; Li, Y.; Zhang, S.; Li, K.; Wang, S.; Tao, P.; Xing, J.; and Lang, C. 2025. Dynamic Dictionary Learning for Remote Sensing Image Segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 1--10

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.