Pith. sign in

REVIEW 4 major objections 6 minor 39 references

Revisiting Cross-Modal Knowledge Distillation: A Disentanglement Approach for RGBD Semantic Segmentation

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a teacher-free, disentanglement-based joint training of RGB and Depth models outperforms teacher/student cross-modal distillation for semantic segmentation while using less compute.

desk verdict Teacher-free CMKD framework with real resource savings, but single-run mIoU with sub-0.3-point margins can't support the 'consistently outperforms' claim. read the letter →

arxiv 2505.24361 v1 pith:GGGHSOR6 submitted 2025-05-30 cs.CV

Roger Ferrod , Cássio F. Dantas , Luigi Di Caro , Dino Ienco This is my paper
classification cs.CV
keywords cross-modalknowledgedistillationRGBDsemanticsegmentationdisentanglementrepresentationlearningcontrastivefeaturemixupdecoupleddataaugmentationteacher-free
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

Cross-modal knowledge distillation (CMKD) traditionally requires a multi-modal teacher network to transfer knowledge into a single-modality student. This paper claims that for RGBD semantic segmentation, the teacher is unnecessary: a teacher-free framework called CroDiNo-KD jointly trains RGB and Depth segmenters, splitting each encoder's features into modality-invariant and modality-specific halves and structuring these halves with orthogonality, contrastive, feature-mixup, and auxiliary losses. Across three datasets (NYUDepth, Potsdam, Mid-Air), CroDiNo-KD reports the best mIoU in both RGB and Depth distillation scenarios, beating recent teacher/student competitors while using less training time and fewer parameters. The authors take this as evidence that the conventional teacher/student paradigm for CMKD should be reconsidered.

What carries the argument

The central mechanism is the forced decomposition of each encoder's embedding map $Z_m \in \mathbb{R}^{B\times h\times w\times F}$ into two halves $Z_m^{\mathrm{inv}}$ and $Z_m^{\mathrm{spc}}$ of $F/2$ channels. Three losses do the structuring: (i) cosine-similarity orthogonality between the two halves at each spatial location (Eq. 5), (ii) an InfoNCE contrastive loss with negative Euclidean distance on spatial-mean-pooled, L2-normalized invariant vectors, pulling invariant embeddings of the same scene across RGB and Depth together and pushing different scenes apart (Eqs. 6–7), and (iii) an auxiliary decoder that independently supervises each half (Eq. 8). On top of this, a feature-mixup blends invariant embeddings across modalities (Eq. 3), and the two modalities are augmented independently, which the paper calls decoupled augmentation. The whole loss is an unweighted sum over RGB and Depth (Eq. 9).

What would settle it

Train CroDiNo-KD on a benchmark such as NYUDepth, then at test time swap the invariant and specific halves $Z_m^{\mathrm{inv}}$ and $Z_m^{\mathrm{spc}}$ before decoding, or substitute random channel partitions; if mIoU does not drop substantially, the disentanglement is not functionally real. A second check: measure the learned cosine similarity between halves – if it is far from zero after training, the orthogonality constraint (Eq. 5) is not being satisfied.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a multi-modal teacher is not needed to distill cross-modal knowledge into single-modality segmentation models. CroDiNo-KD trains two single-modality encoder-decoders (RGB and Depth) simultaneously, dividing each modality's embedding into two halves – an invariant half and a specific half – and enforcing their structure through an unweighted sum of losses: cross-entropy on the full and mixed features, cross-entropy on each half through an auxiliary decoder, cosine orthogonality between the two halves, and InfoNCE contrastive alignment of the invariant halves across modalities. The reported result is consistent improvement over all compared teacher/student frameworks across three benchmarks and both modalities, with a total training-time and parameter budget below every competitor that needs a teacher.

Load-bearing premise

The framework assumes that the combination of orthogonality, contrastive, and auxiliary losses genuinely separates each pixel's features into a modality-invariant and a modality-specific part, so that feature mixup and per-half supervision can treat the two halves as complementary; if the split is not semantically meaningful, the method reduces to an unprincipled sum of losses and its novelty collapses.

Editorial extensions

If this is right

  • Cross-modal distillation for RGBD segmentation can be performed with no teacher and no fusion-network design choices, removing a major source of complexity in current CMKD pipelines.
  • Because the RGB and Depth branches are trained independently and only coupled through losses, each branch can be used at inference time alone, and the approach naturally handles missing-modality deployment.
  • The reported training-cost reduction – about 20.5 GPU-hours with no teacher pre-training versus 36–61 GPU-hours for teacher-based methods, and 73M total parameters versus 178M+ – suggests teacher-free distillation is materially cheaper.
  • The success of the method invites re-evaluating teacher/student paradigms in other multi-modal dense-prediction tasks, such as RGB-thermal or RGB-LiDAR segmentation.

Reading between the lines

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

  • Ablation shows that removing the orthogonality loss ($L_\perp$) barely changes average mIoU (49.22 vs 49.61 full), so the advertised 'disentanglement' may be carried mainly by the contrastive and auxiliary losses rather than by explicit orthogonality; a direct test would be to remove $L_\perp$ and measure whether the invariant/specific split remains semantically meaningful.
  • If the disentanglement is genuinely per-pixel, one could use the invariant half alone as a cross-modal feature for tasks such as retrieval or domain adaptation, or swap halves across modalities to synthesize features of the same scene – the paper does not explore these uses.
  • The decoupled-augmentation property suggests the method could work with unpaired RGB and Depth data or with modalities captured at different resolutions, an extension not tested here but directly supported by the design.
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 / 6 minor

Summary. The paper introduces CroDiNo-KD, a teacher-free cross-modal knowledge distillation framework for RGBD semantic segmentation. It jointly trains two single-modality encoder-decoder models, splitting each encoder's features into modality-invariant and modality-specific halves, and combines cosine orthogonality (Eq. 5), InfoNCE on spatial-mean-pooled invariant features (Eqs. 6-7), feature mixup (Eq. 3), an auxiliary decoder (Eq. 8), and per-modality decoupled augmentation. The method is evaluated on NYU Depth v2, Potsdam, and Mid-Air with mIoU, supported by ablations, sensitivity analysis, training time, and parameter-count comparisons. The paper claims consistent superiority over recent teacher/student CMKD baselines and argues for rethinking the dominant paradigm.

Significance. If the empirical claims are established, the paper offers a useful and economical alternative to teacher-based CMKD, extending disentanglement learning to dense prediction while reducing training cost. The authors evaluate on three heterogeneous benchmarks, compare against several recent methods, and provide resource-efficiency analyses, which are valuable. However, the central claim currently rests on single-run mIoU differences that are often very small, and the semantic validity of the disentanglement is not directly verified. The significance is therefore conditional on stronger empirical support.

major comments (4)
  1. [§4.1, Table 1] The load-bearing sentence 'CroDiNo-KD consistently outperforms all competitors across all benchmarks in both RGB and Depth cross-modal distillation scenarios' is not supported by the evidence. Table 1 reports one run per method, with no error bars, no seed variance, and no significance test. Several margins over the next best competitor are tiny, e.g., Potsdam RGB 76.13 vs. 76.09, Mid-Air RGB 48.37 vs. 48.32, and NYU Depth 37.60 vs. 37.28. These gaps are within typical run-to-run variation for this setup, so the claim requires multi-seed means with standard deviations and, ideally, a paired significance test. This is the main reason the central conclusion is currently unestablished.
  2. [§3.1, Eq. (3) vs. Algorithm 1 line 9] There is a direct inconsistency between the mixup definition in Eq. (3) and the pseudocode. Eq. (3) writes \tilde Z^inv_RGB = λ Z^inv_D + (1-λ) Z^inv_RGB, while Algorithm 1 line 9 writes \tilde Z^inv_m ← λ Z^inv_m + (1-λ) Z^inv_{\bar m}, where \bar m is the complementary modality; for m=RGB this gives the opposite weighting. The paper must specify which convention is actually implemented, since the mixup direction is part of the method and affects reproducibility.
  3. [§3.1, Eqs. (5)-(8) and §4.2, Table 2] The paper never verifies that the orthogonality and contrastive losses actually produce a semantically meaningful invariant/specific split. It only reports end-task mIoU; Table 2 even shows cases where removing a component improves a particular cell (e.g., NYU Depth w/o L_con 37.96 vs. 37.60, Mid-Air RGB w/o L_con 48.46 vs. 48.37, Potsdam Depth w/o L_aux 42.99 vs. 42.78), so the ablation does not confirm that all components are needed. The disentanglement premise is central to the claimed novelty; an independent check (e.g., representation probing, cross-modal retrieval on Z_inv, or measuring the overlap between Z_inv and Z_spc) should be added.
  4. [§4.2, Table 3] The sensitivity analysis selects λ=0.35 after inspecting test-set performance and reports std across λ values, not across training runs. This does not establish robustness to hyperparameter choice, and selecting hyperparameters on the test set can inflate reported performance. The authors should use a validation split for λ selection and report mean±std over multiple seeds for the chosen λ.
minor comments (6)
  1. [Section 2] The sentence 'Taking inspiration from the the standard KD process' contains a duplicated article; please edit.
  2. [Section 4] 'Staring learning rate' should be 'starting learning rate'.
  3. [Section 4.1] 'Postdam' is a typo for 'Potsdam'.
  4. [References] References [1] and [39] appear to refer to the same paper with different years (2022 vs. 2023), and both are used for the KD baselines; please consolidate.
  5. [Abstract and Section 1] 'Source code is available here' does not include a URL; provide a working link.
  6. [Table 5] CroDiNo-KD's main architecture has 68M parameters versus 80M for the baselines; the text should clarify how this is compatible with the statement in Section 4 that all approaches share the same architecture.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical comparison vs external baselines is independent; self-citation [28] is inspirational only.

full rationale

The paper contains no mathematical derivation whose output reduces to an input. Its central claim is empirical: CroDiNo-KD outperforms prior CMKD methods on three public benchmarks. Those baselines (KDv1/2, KD-Net, Masked Dist., ProtoKD, LAD/CAD) are external and are not fitted by CroDiNo-KD's losses, so the comparison is independent. The losses in Eqs. (4)-(9) are optimization objectives, not identities; the assertion that Z_inv and Z_spc encode modality-invariant and modality-specific information is an architectural assumption tested only indirectly through final segmentation accuracy, not a circular definition. Reference [28] (Ienco and Dantas, co-authors of this paper) appears in Related Work as inspiration ('In [28] the authors successfully exploited disentanglement ... Inspired by this pioneering work'), but it is not invoked as a theorem, as evidence of superiority, or as justification for the central result, so it is not load-bearing. The sensitivity analysis of lambda (Table 3) and the single-run mIoU table are experimental-robustness concerns, not circularity: the paper does not claim to predict a quantity that is fit by lambda, and there is no statement that lambda was selected after inspecting test performance. Therefore, no circular step is present; the work is self-contained against external benchmarks.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities, forces, or external constructs. Its main invented elements are architectural, namely the split of encoder features into modality-invariant and modality-specific halves, which is a design choice with no independent falsifiable evidence outside the reported benchmarks.

free parameters (3)
  • mixup lambda = 0.35
    Controls the blending of invariant RGB and Depth features in Eq. 3. Selected via test-set sensitivity analysis in Table 3, not on a separate validation set.
  • InfoNCE temperature tau = 0.07
    Temperature in the contrastive loss in Eq. 7. Set by hand, not ablated, and used in all experiments.
  • encoder output split ratio = F/2 = 512
    Each encoder's representation is split into two halves, Z_inv and Z_spc, to enforce disentanglement. This fixed design choice is not compared against other split ratios.
assumptions (4)
  • domain assumption InfoNCE with in-batch negatives treats all other instances as true negatives regardless of semantic class overlap.
    In Eq. 7, every other sample in the batch is used as a negative. In semantic segmentation, different images frequently contain the same classes, so same-class features may be incorrectly repelled.
  • ad hoc to paper Cosinusoidal orthogonality between Z_inv and Z_spc yields disentangled and complementary representations.
    Eq. 5 minimizes cosine similarity between the two halves, but orthogonality alone does not guarantee that one subspace is modality-invariant and the other modality-specific. The semantic labels are asserted, not measured.
  • domain assumption Feature mixup with linear interpolation between invariant embeddings produces valid training samples for the segmentation decoder.
    Eq. 3 blends Z_inv_RGB and Z_inv_D with a single scalar lambda, assuming that linear paths between deep features are semantically meaningful for both modalities, following Mixup [29] without validation in this setting.
  • domain assumption ImageNet-pretrained ResNet-50 features transfer to depth images after averaging the first-layer RGB weights.
    Section 4, Implementation details, initializes both encoders from ImageNet weights, with the depth encoder's first layer obtained by averaging RGB channels. This assumes depth maps share the same low-level feature statistics as RGB photos.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Cross-Modal Knowledge Distillation: A Disentanglement Approach for RGBD Semantic Segmentation." pith.science (2026). https://pith.science/paper/GGGHSOR6

@misc{pith2026250524361,
  author       = {Pith},
  title        = {Pith review of: Revisiting Cross-Modal Knowledge Distillation: A Disentanglement Approach for RGBD Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GGGHSOR6}},
  note         = {Machine review of arXiv:2505.24361}
}
read the original abstract

Multi-modal RGB and Depth (RGBD) data are predominant in many domains such as robotics, autonomous driving and remote sensing. The combination of these multi-modal data enhances environmental perception by providing 3D spatial context, which is absent in standard RGB images. Although RGBD multi-modal data can be available to train computer vision models, accessing all sensor modalities during the inference stage may be infeasible due to sensor failures or resource constraints, leading to a mismatch between data modalities available during training and inference. Traditional Cross-Modal Knowledge Distillation (CMKD) frameworks, developed to address this task, are typically based on a teacher/student paradigm, where a multi-modal teacher distills knowledge into a single-modality student model. However, these approaches face challenges in teacher architecture choices and distillation process selection, thus limiting their adoption in real-world scenarios. To overcome these issues, we introduce CroDiNo-KD (Cross-Modal Disentanglement: a New Outlook on Knowledge Distillation), a novel cross-modal knowledge distillation framework for RGBD semantic segmentation. Our approach simultaneously learns single-modality RGB and Depth models by exploiting disentanglement representation, contrastive learning and decoupled data augmentation with the aim to structure the internal manifolds of neural network models through interaction and collaboration. We evaluated CroDiNo-KD on three RGBD datasets across diverse domains, considering recent CMKD frameworks as competitors. Our findings illustrate the quality of CroDiNo-KD, and they suggest reconsidering the conventional teacher/student paradigm to distill information from multi-modal data to single-modality neural networks.

Figures

Figures reproduced from arXiv: 2505.24361 by the authors.

Figure 1
Figure 1. Overview of the CroDiNo-KD architecture, composed by two encoder-decoder [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Teacher model architecture used for the competing methods. It consists of [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Example of qualitative results from Potsdam dataset. [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Example of qualitative results from CroDiNo-KD predictions over the MidAir [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 30 canonical work pages

  1. [1]

    The modality focusing hypothesis: Towards under- standing crossmodal knowledge distillation,

    Z. Xue, Z. Gao, S. Ren, and H. Zhao, “The modality focusing hypothesis: Towards under- standing crossmodal knowledge distillation,” in ICLR, 2022

    work page 2022

  2. [2]

    Model compression,

    C. Bucila, R. Caruana, and A. Niculescu-Mizil, “Model compression,” in KDD, 2006

    work page 2006

  3. [3]

    Distilling the knowledge in a neural network,

    G. E. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,”ArXiv, vol. abs/1503.02531, 2015

    arXiv 2015

  4. [4]

    Multi-level logit distillation,

    Y . Jin, J. Wang, and D. Lin, “Multi-level logit distillation,” CVPR, pp. 24276–24285, 2023

    work page 2023

  5. [5]

    Class attention transfer based knowledge distillation,

    Z. Guo, H. Yan, H. Li, and X.L. Lin, “Class attention transfer based knowledge distillation,” CVPR, pp. 11868–11877, 2023

    work page 2023

  6. [6]

    Knowledge distillation from a stronger teacher,

    T. Huang, S. You, F. Wang, C. Qian, and C. Xu, “Knowledge distillation from a stronger teacher,” ArXiv, vol. abs/2205.10536, 2022

    arXiv 2022

  7. [7]

    Lite-mkd: A multi-modal knowledge distillation framework for lightweight few-shot action recognition,

    B. Liu, T. Zheng, P. Zheng, D. Liu, X. Qu, J. Gao, J. Dong, and X. Wang, “Lite-mkd: A multi-modal knowledge distillation framework for lightweight few-shot action recognition,” ACM Multimedia, 2023

    work page 2023

  8. [8]

    Cross-modal distillation for rgb- depth person re-identification,

    F. M. Hafner, A. H. Bhuyian, J. F. P. Kooij, and E. Granger, “Cross-modal distillation for rgb- depth person re-identification,” Comput. Vis. Image Underst., vol. 216, pp. 103352, 2018

    work page 2018

Show all 39 references
  1. [9]

    Knowl- edge distillation from multi-modal to mono-modal segmentation networks,

    M. Hu, M. Maillard, Y . Zhang, T. Ciceri, G. La Barbera, I. Bloch, and P. Gori, “Knowl- edge distillation from multi-modal to mono-modal segmentation networks,” ArXiv, vol. abs/2106.09564, 2020

    2020 arXiv

  2. [10]

    Acnet: Attention based network to exploit comple- mentary features for rgbd semantic segmentation,

    X. Hu, K. Yang, L. Fei, and K. Wang, “Acnet: Attention based network to exploit comple- mentary features for rgbd semantic segmentation,” ICIP, pp. 1440–1444, 2019

    2019

  3. [11]

    Frnet: Frustum-range networks for scalable lidar segmentation,

    X. Xu, L. Kong, H. Shuai, and Q. Liu, “Frnet: Frustum-range networks for scalable lidar segmentation,” ArXiv, vol. abs/2312.04484, 2023

    2023 arXiv

  4. [12]

    Fusenet: Incorporating depth into se- mantic segmentation via fusion-based cnn architecture,

    C. Hazirbas, L. Ma, C. Domokos, and D. Cremers, “Fusenet: Incorporating depth into se- mantic segmentation via fusion-based cnn architecture,” in ACCV, 2016

    2016

  5. [13]

    Indoor semantic segmentation using depth information,

    C. Couprie, C. Farabet, L. Najman, and Y . LeCun, “Indoor semantic segmentation using depth information,” arXiv, 2013

    2013

  6. [14]

    Pixel difference convolutional network for rgb-d semantic segmentation,

    J. Yang, L. Bai, Y . Sun, C. Tian, M. Mao, and G. Wang, “Pixel difference convolutional network for rgb-d semantic segmentation,” IEEE Trans. on Circ. and Sys. for Video Tech. , vol. 34, pp. 1481–1492, 2023

    2023

  7. [15]

    Rdfnet: Rgb-d multi-level residual feature fusion for indoor semantic segmentation,

    S. Lee, S.J. Park, and K. S. Hong, “Rdfnet: Rgb-d multi-level residual feature fusion for indoor semantic segmentation,” ICCV, pp. 4990–4999, 2017

    2017

  8. [16]

    Rednet: Residual encoder-decoder network for indoor rgb-d semantic segmentation,

    J. Jiang, L. Zheng, F. Luo, and Z. Zhang, “Rednet: Residual encoder-decoder network for indoor rgb-d semantic segmentation,” ArXiv, vol. abs/1806.01054, 2018

    2018 arXiv

  9. [17]

    Context-aware interaction network for rgb-t semantic segmenta- tion,

    Y . Lv, Z. Liu, and G. Li, “Context-aware interaction network for rgb-t semantic segmenta- tion,” IEEE Transactions on Multimedia, vol. 26, pp. 6348–6360, 2024

    2024

  10. [18]

    Fuseseg: Semantic segmentation of urban scenes based on rgb and thermal data fusion,

    Y . Sun, W. Zuo, P. Yun, H. Wang, and M. Liu, “Fuseseg: Semantic segmentation of urban scenes based on rgb and thermal data fusion,” IEEE Trans. on Autom. Sci. and Eng. , vol. 18, no. 3, pp. 1000–1011, 2021

    2021

  11. [19]

    Masked generative distillation,

    Z. Yang, Z. Li, M. Shao, D. Shi, Z. Yuan, and C. Yuan, “Masked generative distillation,” in ECCV, 2022. 16 R. Ferrod et al

    2022

  12. [20]

    Prototype knowledge distillation for medical segmentation with missing modality,

    S. Wang, Z. Yan, D. Zhang, H. Wei, Z. Li, and R. Li, “Prototype knowledge distillation for medical segmentation with missing modality,” in ICASSP, 2023

    2023

  13. [21]

    Rethinking knowledge distillation with raw features for semantic segmentation,

    T. Liu, C. Chen, X. Yang, and W. Tan, “Rethinking knowledge distillation with raw features for semantic segmentation,” WACV, pp. 1144–1153, 2024

    2024

  14. [22]

    Disentangled representation learn- ing,

    Xin Wang, Hong Chen, Zihao Wu, Wenwu Zhu, et al., “Disentangled representation learn- ing,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2024

    2024

  15. [23]

    Learning factor- ized multimodal representations,

    Y .H. H. Tsai, P. Pu Liang, A. Zadeh, L.P. Morency, and R. Salakhutdinov, “Learning factor- ized multimodal representations,” ArXiv, vol. abs/1806.06176, 2018

    2018 arXiv

  16. [24]

    Learning disentangled representation for multimodal cross-domain sentiment analysis,

    Y . Zhang, Y . Zhang, W. Guo, X. Cai, and X. Yuan, “Learning disentangled representation for multimodal cross-domain sentiment analysis,” IEEE Trans. on Neural Net. and Learning Sys., vol. 34, no. 10, pp. 7956–7966, 2023

    2023

  17. [25]

    Pre- dict, prevent, and evaluate: Disentangled text-driven image manipulation empowered by pre- trained vision-language model,

    Z. Xu, T. Lin, H. Tang, F. Li, D. He, N. Sebe, R. Timofte, L. Van Gool, and E. Ding, “Pre- dict, prevent, and evaluate: Disentangled text-driven image manipulation empowered by pre- trained vision-language model,” CVPR, pp. 18208–18217, 2021

    2021

  18. [26]

    Towards counterfactual image manipulation via clip,

    Y . Yu, F. Zhan, R. Wu, J. Zhang, S. Lu, M. Cui, X. Xie, X.-S. Hua, and C. Miao, “Towards counterfactual image manipulation via clip,” in ACM Multimedia, 2022

    2022

  19. [27]

    Disentangling visual and written concepts in clip,

    J. Materzy ´nska, A. Torralba, and D. Bau, “Disentangling visual and written concepts in clip,” in CVPR, 2022

    2022

  20. [28]

    Discom-kd: Cross-modal knowledge distillation via disentangle- ment representation and adversarial learning,

    D. Ienco and C.F. Dantas, “Discom-kd: Cross-modal knowledge distillation via disentangle- ment representation and adversarial learning,” in BMVC, 2024

    2024

  21. [29]

    mixup: Beyond empirical risk mini- mization,

    H. Zhang, M. Cissé, Y . Dauphin, and D. Lopez-Paz, “mixup: Beyond empirical risk mini- mization,” ArXiv, vol. abs/1710.09412, 2017

    2017 arXiv

  22. [30]

    Representation learning with contrastive predictive coding,

    A. van den Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” ArXiv, vol. abs/1807.03748, 2018

    2018 arXiv

  23. [31]

    Sigma: Siamese mamba network for multi-modal semantic segmentation,

    Z. Wan, Y . Wang, S. Yong, P. Zhang, S. Stepputtis, K. P. Sycara, and Y . Xie, “Sigma: Siamese mamba network for multi-modal semantic segmentation,”ArXiv, vol. abs/2404.04256, 2024

    2024 arXiv

  24. [32]

    Indoor segmentation and support inference from rgbd images,

    P. Kohli N. Silberman, D. Hoiem and R. Fergus, “Indoor segmentation and support inference from rgbd images,” in ECCV, 2012

    2012

  25. [33]

    Omnivore: A single model for many visual modalities,

    R. Girdhar, M. Singh, N. Ravi, L. van der Maaten, A. Joulin, and I. Misra, “Omnivore: A single model for many visual modalities,” CVPR, pp. 16081–16091, 2022

    2022

  26. [34]

    Urban modelling and semantic labelling benchmark,

    ISPRS, “Urban modelling and semantic labelling benchmark,” 2024, https://www. isprs.org/education/benchmarks/UrbanSemLab/default.aspx

    2024

  27. [35]

    General-purpose multimodal transformer meets remote sensing semantic segmentation,

    N. Kieu, K. Nguyen, S. Sridharan, and C. Fookes, “General-purpose multimodal transformer meets remote sensing semantic segmentation,” ArXiv, vol. abs/2307.03388, 2023

    2023 arXiv

  28. [36]

    Mid-air: A multi-modal dataset for extremely low altitude drone flights,

    M. Fonder and M. Van Droogenbroeck, “Mid-air: A multi-modal dataset for extremely low altitude drone flights,” in CVPRW, June 2019

    2019

  29. [37]

    Unifying distillation and privileged information,

    D. Lopez-Paz, L. Bottou, B. Scholkopf, and V . Naumovich Vapnik, “Unifying distillation and privileged information,” ArXiv, vol. abs/1511.03643, 2015

    2015 arXiv

  30. [38]

    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,” in ECCV, 2018

    2018

  31. [39]

    The modality focusing hypothesis: Towards under- standing crossmodal knowledge distillation,

    Z. Xue, Z. Gao, S. Ren, and H. Zhao, “The modality focusing hypothesis: Towards under- standing crossmodal knowledge distillation,” in ICLR, 2023

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.