Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Hadamard Attention Recurrent Transformer: A Strong Baseline for Stereo Matching Transformer

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that replacing softmax attention with an unbounded positive kernel built from a Hadamard product of query and key norms relieves the low-rank bottleneck in stereo transformers, yielding state-of-the-art disparity accuracy…

desk verdict Solid engineering baseline with real benchmark gains, but the paper's central low-rank-bottleneck story is not supported by its own ablations and the Rank(A)/m metric is ill-posed. read the letter →

arxiv 2501.01023 v5 pith:IOVEZCRF submitted 2025-01-02 cs.CV

Ziyang Chen , Wenting Li , Yongjun Zhang , Yabo Wu , Bingshu Wang , Yong Zhao , C.L. Philip Chen This is my paper · ORCID
classification cs.CV
keywords Hadamardattentionstereomatchinglineartransformerlow-rankbottleneckdensekernelmulti-kernelinteractionrecurrentreflectivesurface
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

HART is a stereo-matching transformer built on Hadamard-product attention that runs in linear time rather than quadratic time. The paper's central claim is that the Dense Attention Kernel (DAK), which maps attention weights to the unbounded interval $(0, +\infty)$ instead of softmax's bounded range, relieves the low-rank bottleneck of self-attention and reduces feature collinearity, letting the model match points in ill-posed regions such as reflective surfaces. To compensate for the Hadamard product's lack of spatial and channel interaction, the Multi-Kernel & Order Interaction (MKOI) module mixes features across different kernel sizes and channel groups. The paper reports an end-point error of 0.42 pixels on Scene Flow, first place on the KITTI 2012 reflective leaderboard at the time of submission, and zero-shot generalization to Middlebury and Driving Stereo with lower errors than prior stereo transformers. If correct, the method offers a scalable attention design that could serve as a new baseline for stereo transformer models.

What carries the argument

The load-bearing object is the Hadamard Product Self-Attention (HPSA) mechanism, whose attention matrix is computed as the element-wise (Hadamard) product of the squared norms of query and key, giving $O(n)$ complexity instead of $O(n^2)$. Dense Attention Kernel (DAK) is the activation that keeps every entry positive and unbounded, while Multi-Kernel & Order Interaction (MKOI) splits the channels into groups processed with kernels of sizes $1\times 1$, $3\times 3$, $5\times 5$, and $7\times 7$ to restore spatial and channel mixing that pure Hadamard products lack. Together these components replace softmax attention in the encoder, and the decoder follows an iterative update scheme with an LSTM-based refinement module inspired by recurrent stereo networks.

What would settle it

Train the same HART architecture with DAK replaced by another positive unbounded activation such as $\mathrm{ELU}(x)+1$ or Softplus, keeping all other settings identical. If end-point error on Scene Flow and reflective-region errors stay essentially the same while $\mathrm{Rank}(A)/m$ changes, then the low-rank-bottleneck story is not what carries the performance; if $\mathrm{Rank}(A)/m$ rises but accuracy does not, the rank metric is not predictive.

Watch

Extended reading notes

Core claim

The paper's core discovery is that replacing the matmul-plus-softmax attention in stereo transformers with an attention matrix $A = \|Q\|^2 \odot \|K\|^2$, followed by a dense kernel $\mathrm{DAK}(A) = A+1$ for $A\ge 0$ and $e^{A}$ for $A<0$, produces a positive attention matrix that keeps all parameters active and improves disparity accuracy in difficult regions. The authors argue that removing the upper bound on attention weights lowers feature similarity and thereby counteracts the low-rank bottleneck identified for standard multi-head attention. The empirical anchor is an ablation study showing that the ratio $\mathrm{Rank}(A)/m$ rises from 78.1% with vanilla self-attention to 94.5% with DAK plus MKOI, while end-point error on Scene Flow drops from 0.44 to 0.42 pixels and inference time falls from 0.71 seconds to 0.36 seconds.

Load-bearing premise

The argument assumes that the rank of the attention matrix, as measured by the paper's Rank(A)/m metric, is the real driver of matching accuracy rather than a side effect of DAK being a better activation function.

Editorial extensions

If this is right

  • Attention in stereo transformers can be made linear-time without sacrificing accuracy, removing a computational barrier to high-resolution and real-time stereo matching.
  • DAK's unbounded positive kernel could be dropped into other vision transformers that suffer from low-rank attention, potentially improving accuracy in ill-posed regions beyond stereo.
  • The rank-based diagnostic $\mathrm{Rank}(A)/m$ gives practitioners a cheap way to compare attention mechanisms by representational richness, not just by downstream error.
  • Improved matching on reflective surfaces suggests safer disparity estimates for autonomous driving on glass, wet roadways, and other specular scenes.

Reading between the lines

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

  • Editorial inference: the paper's argument implies that any activation keeping attention weights positive and unbounded, such as $\mathrm{ELU}(x)+1$ or Softplus, might capture part of the benefit; the unique contribution of DAK's piecewise form would need to be isolated by a controlled experiment the paper does not run.
  • Editorial inference: the $\mathrm{Rank}(A)/m$ metric is a self-defined proxy; a stronger test would compute the numerical rank or effective rank with a tolerance and correlate rank increases with accuracy gains per region type rather than across a whole benchmark.
  • Editorial inference: because MKOI adds back local convolutions, HPSA is best understood as a convolution-attention hybrid, and the paper's claim of global interaction may hold only at the level of the full multi-scale encoder rather than within a single attention step.
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. HART proposes a stereo matching transformer built on Hadamard-product self-attention (HPSA), with two main components: the Dense Attention Kernel (DAK), which replaces softmax with a positive-valued activation to supposedly overcome a low-rank bottleneck, and the Multi-Kernel & Order Interaction (MKOI) module, which adds spatial and channel interactions via channel splitting and multi-scale convolutions. The encoder is paired with an IGEV-style recurrent decoder. The paper reports SOTA or near-SOTA results on Scene Flow, KITTI 2012/2015, Middlebury, and zero-shot Driving Stereo, including first place on the KITTI 2012 reflective benchmark at the time of submission. The central causal claim is that DAK raises the rank of the attention matrix, reduces feature collinearity, and thereby improves matching in ill-posed regions such as reflections.

Significance. If the causal story were supported, HART would be a notable contribution: it offers linear-complexity attention for stereo transformers, strong empirical results across multiple benchmarks, zero-shot generalization experiments, and an extension replacing DLNR's attention module. The manuscript ships code and follows standard training protocols, which is a strength. However, the paper's load-bearing theoretical claim—that DAK alleviates a low-rank bottleneck and that this mechanism drives the accuracy gains—is not established by the evidence presented. The ablation metric used to support the claim is not well-defined across architectures, and the ablations do not show that rank tracks accuracy. The empirical contribution is real, but the causal narrative needs either rigorous support or substantial softening.

major comments (4)
  1. [§4.4, Table 8] The Rank(A)/m metric is not a well-defined cross-architecture comparison. In Experiment 2, the attention matrix for vanilla SA is n×n, while in Experiments 3–6 the Hadamard-product attention matrix is c×n; comparing rank ratios of matrices with different shapes and different semantic roles is not meaningful. In addition, the known low-rank bound for softmax attention (rank at most the head dimension) does not transfer to the c×n elementwise product used here. To support the claim that DAK alleviates a low-rank bottleneck, the authors need to define a metric on the actual attention operator used in each variant and verify that higher rank is causally tied to lower EPE.
  2. [§4.4, Table 8] The ablation results contradict the rank-to-accuracy link. Row 4 (DAK without MKOI) has Rank(A)/m=93.0 but EPE=0.46, which is worse than row 5 (Softmax with MKOI) at Rank(A)/m=81.8 and EPE=0.43; row 5 achieves better accuracy despite a lower rank. Within the same MKOI setting, replacing Softmax with DAK changes EPE only from 0.43 to 0.42. This pattern indicates that MKOI and the overall architecture, rather than DAK's rank effect, are the main drivers of the reported gains. The paper should provide a controlled experiment that isolates DAK's effect on rank and matching accuracy, or explicitly drop the rank-based causal claim.
  3. [§3.2.2, Eq. (6)] Equation (6) is not a valid derivation of a DAK-specific benefit. Since DAK(A)=1+ELU(A), Eq. (5) is simply a residual ELU, and Eq. (6) is the standard chain rule for a residual connection; the expression inside MKOI mixes a scalar with a derivative of a sum and is dimensionally unclear. The paper presents this as a theoretical justification for DAK, but it does not establish that DAK removes a rank bottleneck. Either provide a rigorous argument linking DAK to rank improvement or temper the theoretical framing.
  4. [§3.2.1 and §3.2.2] The parameter-count argument '2nc/d << n^2' does not prove a low-rank bottleneck. The rank of the attention matrix is bounded by its dimensions and the algebraic structure of Q and K, not by the number of parameters in the query and key matrices. The paper should state precisely which matrix's rank it analyzes, before or after DAK/Softmax, and what exact rank bound DAK is claimed to escape; otherwise the low-rank-bottleneck narrative is unsupported.
minor comments (6)
  1. [§3.2, heading] The heading 'Caculation of Attention Matrix' contains a typo; it should read 'Calculation of Attention Matrix'.
  2. [Abstract and Introduction] There are duplicated words in the text, e.g., 'and and secured a position' in the Introduction; the manuscript needs a careful proofread.
  3. [§3.2.2 and §4.4] It is unclear whether the Rank(A)/m metric in Table 8 is computed on A before or after applying DAK/Softmax and MKOI; this should be stated explicitly, along with the number of inference steps and samples used to compute the average.
  4. [§3.2.2] The phrase 'the derivative properties of DAK determine its ability to further widen the attentional gap' is vague; please specify the derivative inequality and how it redirects attention.
  5. [§3.3, Eq. (7)] The notation in Eq. (7) is dense and the kernel-size indexing s=2m+3 is not defined in the text around the equation; adding a short explanation would improve readability.
  6. [§4.2] The claim 'HART ranked 1st at the time of submission' is time-dependent; the paper should note that leaderboard positions may change and the evaluation is as of the submission date.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HART's benchmark claims are externally validated, and the DAK/rank-metric discussion is definitional rather than a fitted prediction.

full rationale

HART's derivation chain is self-contained against external benchmarks. The central empirical claims (0.42 EPE on Scene Flow, KITTI 2012 reflective rank 1, Middlebury and Driving Stereo zero-shot numbers) are benchmark results obtained with standard training and fine-tuning protocols, not quantities fitted from the claimed predictions. DAK is defined by Eq. 4 and is algebraically identical to 1+ELU via Eq. 5; this is a definitional identity, not a prediction derived from the low-rank hypothesis, so it does not make the argument circular. The Rank(A)/m metric in Tab. 8 is a self-defined diagnostic: DAK's positivity converts zero elements into nonzero ones and therefore tends to raise this rank measure by construction, but the paper does not use this metric as the target outcome — EPE and D1 are reported separately, and the ablation table does not show a monotonic rank-to-accuracy correspondence (e.g., row 4 has higher rank than row 2 but worse EPE, while row 5 has lower rank than row 4 but better EPE). That weakens the low-rank-bottleneck explanation as a causal story, but it is a correctness or validity concern, not a circular reduction. Self-citations such as MoCha-Stereo [6] and refs. [13,29] are used as SOTA baselines or background material and are not load-bearing for HART's central claim. The paper's own stated limitation (HART is not real-time) is a limitation, not a circular step. No circularity is established.

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

The central design leans on the low-rank bottleneck premise and the rank proxy; no new physical entities are introduced.

free parameters (3)
  • gamma (loss decay weight) = 0.9
    Set by hand in Eq. 1; standard exponential weighting, not fitted to benchmark results.
  • Number of recurrent iterations (evaluation) = 32
    Standard choice following IGEV-Stereo; affects accuracy and speed trade-off.
  • Max disparity of correlation encoding volume = 192
    Hyperparameter for correlation volume, chosen for dataset disparity range.
assumptions (4)
  • domain assumption Low-rank bottleneck in vanilla self-attention is the main cause of poor matching in ill-posed regions
    Invoked in Sec. 1 and 3.2.1 based on [9], but not established specifically for stereo matching.
  • domain assumption The rank of the Hadamard-product attention matrix A is a valid proxy for model expressiveness
    Introduced in Sec. 4.4 as Rank(A)/m metric; no independent validation that higher rank improves disparity accuracy.
  • domain assumption Hadamard product attention combined with MKOI can capture long-range spatial interactions needed for stereo matching
    The architecture's core premise; large kernels in MKOI provide limited receptive fields, not global attention.
  • standard math Standard gradient chain rule (Eq. 6) applied to residual connections is valid
    Used in Sec. 3.2.2 to support gradient flow; this is well-known background.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hadamard Attention Recurrent Transformer: A Strong Baseline for Stereo Matching Transformer." pith.science (2026). https://pith.science/paper/IOVEZCRF

@misc{pith2026250101023,
  author       = {Pith},
  title        = {Pith review of: Hadamard Attention Recurrent Transformer: A Strong Baseline for Stereo Matching Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IOVEZCRF}},
  note         = {Machine review of arXiv:2501.01023}
}
abstract

Constrained by the low-rank bottleneck inherent in attention mechanisms, current stereo matching transformers suffer from limited nonlinear expressivity, which renders their feature representations sensitive to challenging conditions such as reflections. To overcome this difficulty, we present the Hadamard Attention Recurrent Stereo Transformer (HART). HART includes a novel attention mechanism that incorporates the following components: 1) The Dense Attention Kernel (DAK) maps the attention weight distribution into a high-dimensional space over (0, +$\infty$). By removing the upper bound constraint on attention weights, DAK enables more flexible modeling of complex feature interactions. This reduces feature collinearity. 2) The Multi Kernel & Order Interaction (MKOI) module extends the attention mechanism by unifying semantic and spatial knowledge learning. This integration improves the ability of HART to learn features in binocular images. Experimental results demonstrate the effectiveness of our HART. In reflective area, HART ranked 1st on the KITTI 2012 benchmark among all published methods at the time of submission. Code is available at https://github.com/ZYangChen/HART.

Figures

Figures reproduced from arXiv: 2501.01023 by the authors.

Figure 1
Figure 1. Motivation. The ability to effectively and efficiently express in ill-posed regions, i.e. reflective area, is a crucial aspect of the stereo matching application. The low-rank bottleneck and the quadratic time complexity present significant challenges for stereo transformers in meeting the aforementioned needs. Our objective is to address these issues through the implementation of HART. Since the development of deep… view at source ↗
Figure 2
Figure 2. The overall architecture of our HART. We construct multi-scale transformer blocks to encode features. The attention and forward [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Overview of Multi Kernel & Order Interaction (MKOI). The left box shows the splitting and convolution of the Attention matrix A, and the right box shows the splitting and convolution of the Value matrix V. It is worth noting that the letters in the brackets represent the number of channels. Lines of the same color represent the corresponding Hadamard product. global semantic distributions. Compared to applying large… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Visualisation on the KITTI dataset [32]. We conducted comparisons with existing SOTA methods [5, 42, 50]. In the first image, HART [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Visualisation on the Middlebury dataset [34]. Disparity and error maps at 0.5 px threshold are included. In the error maps, black [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Zero-shot evaluation without fine-tuning on Middlebury. All results visualised here are trained on Scene Flow. The odd-numbered [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Visualization of the effects using DAK and SoftMax as kernel functions is presented. The figure shows a visualization of the model trained on the Scene Flow Train dataset, evaluated on a sample scene from the Scene Flow Test dataset. As highlighted in the zoomed-in reg…
Figure 8
Figure 8. Figure 8: Visualisation of the zero-shot performance comparison between [16] and our HART-DLNR. The results show that our HPSA can achieve [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Motif Channel Opened in a White-Box: Stereo Matching via Motif Correlation Graph

    cs.CV 2024-11 conditional novelty 5.0 of 10

    MoCha-V2 combines a parameter-free Motif Correlation Graph over wavelet-domain features with iterative refinement to achieve top stereo matching accuracy on multiple public benchmarks.

  2. WHTMix: Efficient Stereo Depth Estimation via Walsh-Hadamard Token Mixing

    cs.CV 2026-07 conditional novelty 4.0 of 10

    A fixed Walsh-Hadamard token mixer can replace the joint self-attention in a stereo transformer at parity accuracy on synthetic data, with 2.46x less compute and 2.65x lower latency, governed by the token-to-channel ratio.

Reference graph

Works this paper leans on

55 extracted references · 51 canonical work pages · cited by 2 Pith papers

  1. [1]

    Surface depth estimation from multi-view stereo satellite images with distribution contrast network

    Ziyang Chen, Wenting Li, Zhongwei Cui, and Yongjun Zhang. Surface depth estimation from multi-view stereo satellite images with distribution contrast network. IEEE J. Sel. Topics Appl. Earth Observ.Remote Sens., 2024

    work page 2024

  2. [2]

    Prior information guided semi-supervised semantic segmentation of remote sensing images

    Xiaorong Gan, Wenting Li, Yongjun Zhang, Wei Long, Yujie Lu, and Ziyang Chen. Prior information guided semi-supervised semantic segmentation of remote sensing images. IEEE Trans. Geoscience Remote Sensing, 2025

    work page 2025

  3. [3]

    Rethinking low-light knowledge for pedestrian detection in nighttime conditions

    Ziyang Chen, He Yao, Wenting Li, Wei Long, and Yongjun Zhang. Rethinking low-light knowledge for pedestrian detection in nighttime conditions. Engineering Applications of Artificial Intelligence, 2025

    work page 2025

  4. [4]

    Raft-stereo: Multilevel recurrent field transforms for stereo matching

    Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In Int. Conf. 3D Vis., pages 218–227. IEEE, 2021

    work page 2021

  5. [5]

    Iterative geometry encoding volume for stereo matching

    Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21919–21928, 2023

    work page 2023

  6. [6]

    Mocha-stereo: Motif channel attention network for stereo matching

    Ziyang Chen, Wei Long, He Yao, Yongjun Zhang, Bingshu Wang, Yongbin Qin, and Jia Wu. Mocha-stereo: Motif channel attention network for stereo matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 27768–27777, 2024

    work page 2024

  7. [7]

    Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers

    Zhaoshuo Li, Xingtong Liu, Nathan Drenkow, Andy Ding, Francis X Creighton, Russell H Taylor, and Mathias Unberath. Revisiting stereo depth estimation from a sequence-to-sequence perspective with transformers. In Int. Conf. Comput. Vis., pages 6197–6206, 2021

    work page 2021

  8. [8]

    Elfnet: Evidential local-global fusion for stereo matching

    Jieming Lou, Weide Liu, Zhuo Chen, Fayao Liu, and Jun Cheng. Elfnet: Evidential local-global fusion for stereo matching. In Int. Conf. Comput. Vis., pages 17784–17793, 2023

    work page 2023

Show all 55 references
  1. [9]

    Low-rank bottleneck in multi-head attention models

    Srinadh Bhojanapalli, Chulhee Yun, Ankit Singh Rawat, Sashank Reddi, and Sanjiv Kumar. Low-rank bottleneck in multi-head attention models. In Int. Conf. Mach. Learn., pages 864–873. PMLR, 2020

    2020

  2. [10]

    Group-wise correlation stereo network

    Xiaoyang Guo, Kai Yang, Wukui Yang, Xiaogang Wang, and Hongsheng Li. Group-wise correlation stereo network. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3273–3282, 2019

    2019

  3. [11]

    Feature distribution normalization network for multi-view stereo

    Ziyang Chen, Yang Zhao, Junling He, Yujie Lu, Zhongwei Cui, Wenting Li, and Yongjun Zhang. Feature distribution normalization network for multi-view stereo. The Vis. Comput., pages 1–13, 2024

    2024

  4. [12]

    Practical stereo matching via cascaded recurrent network with adaptive correlation

    Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 16263–16272, 2022

    2022

  5. [13]

    Leveraging negative correlation for full-range self-attention in vision transformers

    Wei Long, Ziyang Chen, Wenting Li, Yongjun Zhang, He Yao, Jiaxin Peng, and Zhongwei Cui. Leveraging negative correlation for full-range self-attention in vision transformers. Pattern Recognit., page 111899, 2026

    2026

  6. [14]

    Croco v2: Improved cross-view completion pre-training for stereo matching and optical flow

    Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Br ´egier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and J´erˆome Revaud. Croco v2: Improved cross-view completion pre-training for stereo matching and optical flow. In Int. Conf....

    2023

  7. [15]

    A frequency-domain dynamic amplitude filtering method for single-image dehazing with harmony enhancement

    Yabo Wu, Yongjun Zhang, Ziyang Chen, and Yong Zhao. A frequency-domain dynamic amplitude filtering method for single-image dehazing with harmony enhancement. Expert Systems with Applications, page 128542, 2025

    2025

  8. [16]

    High-frequency stereo matching network

    Haoliang Zhao, Huizhou Zhou, Yongjun Zhang, Jie Chen, Yitong Yang, and Yong Zhao. High-frequency stereo matching network. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1327–1336, 2023

    2023

  9. [17]

    Retinastereo: Dynamic- volume stereo matching network

    Xiaoyan Liao, Haoliang Zhao, Fan Yang, Kwokching Cheung, Jun Jiang, Yong Zhao, Jie Chen, and Xinan Wang. Retinastereo: Dynamic- volume stereo matching network. In IEEE Int. Conf. Acoustics, Speech Sign. Process., pages 1–5. IEEE, 2025

    2025

  10. [18]

    Restormer: E fficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: E fficient transformer for high-resolution image restoration. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5728–5739, 2022. 20

    2022

  11. [19]

    Catnet: Coordinate-aware transformer for all-in-one image restoration

    Junling He, Yang Zhao, Wenting Li, Ziyang Chen, Yao Xiao, Bingshu Wang, and Yongjun Zhang. Catnet: Coordinate-aware transformer for all-in-one image restoration. Expert Systems with Applications, 296:128896, 2026

    2026

  12. [20]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Adv. Neural Inform. Process. Syst., 30, 2017

    2017

  13. [21]

    Linformer: Self-attention with linear complexity

    Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768, 2020

    2006 arXiv

  14. [22]

    E fficient attention: Attention with linear complexities

    Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. E fficient attention: Attention with linear complexities. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3531–3539, 2021

    2021

  15. [23]

    Attention is not all you need: Pure attention loses rank doubly exponentially with depth

    Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In Int. Conf. Mach. Learn., pages 2793–2803. PMLR, 2021

    2021

  16. [24]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770–778, 2016

    2016

  17. [25]

    Visualizing and understanding convolutional networks

    Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In Eur. Conf. Comput. Vis., pages 818–833. Springer, 2014

    2014

  18. [26]

    Large kernel matters–improve semantic segmentation by global convo- lutional network

    Chao Peng, Xiangyu Zhang, Gang Yu, Guiming Luo, and Jian Sun. Large kernel matters–improve semantic segmentation by global convo- lutional network. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4353–4361, 2017

    2017

  19. [27]

    Hornet: Efficient high-order spatial interactions with recursive gated convolutions

    Yongming Rao, Wenliang Zhao, Yansong Tang, Jie Zhou, Ser Nam Lim, and Jiwen Lu. Hornet: Efficient high-order spatial interactions with recursive gated convolutions. Adv. Neural Inform. Process. Syst., 35:10353–10366, 2022

    2022

  20. [28]

    Layer normalization

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geo ffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016

    2016 arXiv

  21. [29]

    Motif channel opened in a white-box: Stereo matching via motif correlation graph

    Ziyang Chen, Yongjun Zhang, Wenting Li, Bingshu Wang, Yong Zhao, and CL Chen. Motif channel opened in a white-box: Stereo matching via motif correlation graph. arXiv preprint arXiv:2411.12426, 2024

    2024 arXiv

  22. [30]

    E fficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. E fficientnet: Rethinking model scaling for convolutional neural networks. In Int. Conf. Mach. Learn. , pages 6105–6114. PMLR, 2019

    2019

  23. [31]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3354–3361. IEEE, 2012

    2012

  24. [32]

    Object scene flow for autonomous vehicles

    Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. InIEEE Conf. Comput. Vis. Pattern Recog., pages 3061–3070, 2015

    2015

  25. [33]

    A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation

    Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4040–4048, 2016

    2016

  26. [34]

    High-resolution stereo datasets with subpixel-accurate ground truth

    Daniel Scharstein, Heiko Hirschm ¨uller, York Kitajima, Greg Krathwohl, Nera Neˇsi´c, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In Pattern Recognit. German Conf., pages 31–42. Springer, 2014

    2014

  27. [35]

    Tartanair: A dataset to push the limits of visual slam

    Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In Int. Conf. Intell. Robots Syst., pages 4909–4916. IEEE, 2020

    2020

  28. [36]

    Falling things: A synthetic dataset for 3d object detection and pose estimation

    Jonathan Tremblay, Thang To, and Stan Birchfield. Falling things: A synthetic dataset for 3d object detection and pose estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2038–2041, 2018

    2018

  29. [37]

    Instereo2k: a large real dataset for stereo matching in indoor scenes

    Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Sci. China Inf. Sci., 63:1–11, 2020

    2020

  30. [38]

    Hierarchical deep stereo matching on high-resolution images

    Gengshan Yang, Joshua Manela, Michael Happold, and Deva Ramanan. Hierarchical deep stereo matching on high-resolution images. In 21 IEEE Conf. Comput. Vis. Pattern Recog., pages 5515–5524, 2019

    2019

  31. [39]

    End-to-end learning of geometry and context for deep stereo regression

    Alex Kendall, Hayk Martirosyan, Saumitro Dasgupta, Peter Henry, Ryan Kennedy, Abraham Bachrach, and Adam Bry. End-to-end learning of geometry and context for deep stereo regression. In Int. Conf. Comput. Vis., pages 66–75, 2017

    2017

  32. [40]

    Ga-net: Guided aggregation net for end-to-end stereo matching

    Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 185–194, 2019

    2019

  33. [41]

    Aanet: Adaptive aggregation network for efficient stereo matching

    Haofei Xu and Juyong Zhang. Aanet: Adaptive aggregation network for efficient stereo matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1959–1968, 2020

    1959

  34. [42]

    Accurate and efficient stereo matching via attention concatenation volume

    Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Trans. Pattern Anal. Mach. Intell., 2023

    2023

  35. [43]

    Selective-stereo: Adaptive frequency information selection for stereo matching

    Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. Selective-stereo: Adaptive frequency information selection for stereo matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19701–19710, 2024

    2024

  36. [44]

    Global occlusion-aware transformer for robust stereo matching

    Zihua Liu, Yizhou Li, and Masatoshi Okutomi. Global occlusion-aware transformer for robust stereo matching. In IEEE Winter Conf. Appl. Comput. Vis., pages 3535–3544, 2024

    2024

  37. [45]

    P3snet: Parallel pyramid pooling stereo network

    Alper Emlek and Murat Peker. P3snet: Parallel pyramid pooling stereo network. IEEE Trans. Intell. Transp. Syst. , 24(10):10433–10444, 2023

    2023

  38. [46]

    Digging into uncertainty-based pseudo-label for robust stereo matching

    Zhelun Shen, Xibin Song, Yuchao Dai, Dingfu Zhou, Zhibo Rao, and Liangjun Zhang. Digging into uncertainty-based pseudo-label for robust stereo matching. IEEE Trans. Pattern Anal. Mach. Intell., 2023

    2023

  39. [47]

    Adaptive multi-modal cross-entropy loss for stereo matching

    Peng Xu, Zhiyu Xiang, Chengyu Qiao, Jingyun Fu, and Tianyu Pu. Adaptive multi-modal cross-entropy loss for stereo matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5135–5144, 2024

    2024

  40. [48]

    Los: Local structure-guided stereo matching

    Kunhong Li, Longguang Wang, Ye Zhang, Kaiwen Xue, Shunbo Zhou, and Yulan Guo. Los: Local structure-guided stereo matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19746–19756, 2024

    2024

  41. [49]

    Adstereo: Efficient stereo matching with adaptive downsampling and disparity alignment

    Yun Wang, Kunhong Li, Longguang Wang, Junjie Hu, Dapeng Oliver Wu, and Yulan Guo. Adstereo: Efficient stereo matching with adaptive downsampling and disparity alignment. IEEE Trans. Image Process., 2025

    2025

  42. [50]

    Any-stereo: Arbitrary scale disparity estimation for iterative stereo matching

    Zhaohuai Liang and Changhe Li. Any-stereo: Arbitrary scale disparity estimation for iterative stereo matching. In AAAI, volume 38, pages 3333–3341, 2024

    2024

  43. [51]

    Robust synthetic-to-real transfer for stereo matching

    Jiawei Zhang, Jiahe Li, Lei Huang, Xiaohan Yu, Lin Gu, Jin Zheng, and Xiao Bai. Robust synthetic-to-real transfer for stereo matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 20247–20257, 2024

    2024

  44. [52]

    A transformer-based architecture for high-resolution stereo matching.IEEE Trans

    Di Jia, Peng Cai, Qian Wang, and Ninghua Yang. A transformer-based architecture for high-resolution stereo matching.IEEE Trans. Comput. Imaging, 10:83–92, 2024

    2024

  45. [53]

    Context-enhanced stereo transformer

    Weiyu Guo, Zhaoshuo Li, Yongkui Yang, Zheng Wang, Russell H Taylor, Mathias Unberath, Alan Yuille, and Yingwei Li. Context-enhanced stereo transformer. In Eur. Conf. Comput. Vis., pages 263–279. Springer, 2022

    2022

  46. [54]

    Cfnet: Cascade and fused cost volume for robust stereo matching

    Zhelun Shen, Yuchao Dai, and Zhibo Rao. Cfnet: Cascade and fused cost volume for robust stereo matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 13906–13915, 2021

    2021

  47. [55]

    Drivingstereo: A large-scale dataset for stereo matching in autonomous driving scenarios

    Guorun Yang, Xiao Song, Chaoqin Huang, Zhidong Deng, Jianping Shi, and Bolei Zhou. Drivingstereo: A large-scale dataset for stereo matching in autonomous driving scenarios. In IEEE Conf. Comput. Vis. Pattern Recog., pages 899–908, 2019. 22

    2019

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.