Pith. sign in

REVIEW 4 major objections 5 minor 32 references

Rethinking Encoder-Decoder Flow Through Shared Structures

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that feeding every decoder block two shared tensors — a feature bank and a sampling bank — improves monocular depth estimation for ViT and RepViT encoders while adding under 1% parameters for ViT-based models.

desk verdict A genuinely cheap and plausible decoder mechanism with a fair internal comparison, but the natural-image headline numbers are teacher-imitation scores, so the paper's unqualified improvement claim is not yet grounded. read the letter →

arxiv 2501.14535 v1 pith:RWASWMTI submitted 2025-01-24 cs.CV cs.LG

Frederik Laboyrie , Mehmet Kerim Yucel , Albert Saa-Garriga This is my paper · ORCID
classification cs.CVcs.LG
keywords monoculardepthestimationtransformerencodersdecoderarchitecturesharedfeaturebanksguidedsamplingdynamicupsamplingdenseprediction
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

Decoder blocks in dense prediction typically consume one intermediate feature map plus the previous block's output, so no block sees the whole encoder. This paper proposes two shared tensors, a feature bank and a sampling bank, generated once from all encoder feature maps and consumed by every decoder block. The authors report that this change improves monocular depth estimation on natural and synthetic images for ViT and RepViT encoders, with parameter increases of under 1% for ViT-based models. The sampling bank's guided dynamic upsampling is credited with most of the qualitative gains, such as sharper edges and fewer ghosting artifacts. If the result holds, banks are a cheap generic upgrade to encoder-decoder architectures for dense prediction.

What carries the argument

The shared bank generator $Z$ is a pair of residual pointwise convolutional blocks that take all encoder intermediate feature maps and output $B_{\text{sample}}$ (sampling bank) and $B_{\text{feat}}$ (feature bank). Each decoder block receives both tensors as extra inputs. The feature bank interacts with a block's fused map $X$ through $B' = \operatorname{conv}(\operatorname{concat}(B, X))$ and reweights it as $X' = X \times B'$. The sampling bank is used by a guided sampling function $\operatorname{GS}_{\uparrow/\downarrow}(X_{\text{in}}, X_{\text{ref}})$, a reconfiguration of DySample in which a reference tensor generates offset residuals directly at the target resolution, allowing non-integer sampling factors and downsampling; within a block the operation is $O = \operatorname{GS}_\uparrow(X, \operatorname{GS}_\downarrow(B_{\text{sample}}, X))$.

What would settle it

Retrain the same four encoder-decoder pairs with and without banks on the large mixed dataset, then evaluate on the original ground-truth depth maps of NYUv2, Mannequin Challenge, and IRS instead of pseudo-labels; if the bank models fail to beat the no-bank models on $\delta_1$ and AbsRel, the central claim of improved depth estimation collapses.

Watch

Extended reading notes

Core claim

The central claim is that decoding can be improved by giving every decoder block access to global shared context. Formally, the per-block equation $D_i(O_{E_i}, O_{D_{i-1}})$ is replaced by $D_i(O_{E_i}, O_{D_{i-1}}, B_{\text{sample}}, B_{\text{feat}})$, where the two bank tensors are produced once by a generator $Z(E)$ from all encoder feature maps. The feature bank reweights each block's fused features through concatenation, joint convolution, and element-wise multiplication; the sampling bank acts as a reference tensor for a guided sampling operation that supports both upsampling and downsampling, including non-integer factors. In their experiments, banks increase $\delta_1$ and reduce AbsRel for ViT-S, ViT-B, RepViT m1.1, and RepViT m2.3 on NYUv2, Mannequin Challenge, and IRS after training on a large mixed dataset, and on HyperSim under synthetic ground truth. The price is under 1% extra parameters for ViT-based models and roughly 3–5% for RepViT, with GFLOPs rising about 4–6% for ViT and 3% for RepViT.

Load-bearing premise

The reported accuracy gains on NYUv2, Mannequin Challenge, and IRS are measured against DepthAnythingV2 pseudo-labels rather than the datasets' ground truth, so if those pseudo-labels share errors with the encoder being evaluated, the measured improvement may not reflect true depth accuracy.

Editorial extensions

If this is right

  • Any encoder that emits multiple intermediate feature maps can adopt banks without changing the encoder or the training loss, so the mechanism should transfer to other dense prediction tasks such as semantic segmentation.
  • A bank-enabled ViT-S reaches $\delta_1=0.819$ on NYUv2, close to a plain ViT-B's 0.826, while adding under 1% parameters to the smaller model.
  • The guided sampling mechanism, even without dynamic downsampling of the bank, is the main source of qualitative improvement: fewer ghosting artifacts, sharper edges, and recovery of block artifacts.
  • On the smaller HyperSim dataset, bank gains are more consistent for the convolutional RepViT encoders than for ViTs, suggesting the benefit interacts with data scale and architecture.

Reading between the lines

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

  • Inference: If the pseudo-label evaluation bias is small, the bank mechanism is decoupled from the loss and should also improve metric-depth heads such as bins-based decoders, which the paper does not test.
  • Inference: The bank generator uses only residual pointwise convolutions; the paper notes generation and interaction 'could be much heavier,' so scaling bank capacity or making the banks hierarchical is a natural next experiment that could enlarge the reported gains.
  • Inference: Because the banks summarize the entire encoder, they could be used to condition decoder blocks on global scene context in tasks beyond depth, such as surface-normal or segmentation heads, without retraining the encoder.
  • Inference: The guided sampling formulation's support for non-integer factors and downsampling could simplify handling of odd feature-map dimensions in other transformer backbones, not just the ViT configurations tested here.
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 a modification to encoder-decoder architectures for monocular depth estimation: two globally shared tensors, a feature bank and a sampling bank, are generated once from all intermediate encoder feature maps and are consumed by every decoder block through element-wise feature reweighting and a guided DySample-style upsampling/downsampling procedure. The method is evaluated with ViT-S, ViT-B, and RepViT encoders on a large-scale mixed natural/synthetic training set, with natural-image metrics computed on NYUv2, Mannequin Challenge, and IRS, and a separate synthetic ground-truth evaluation on HyperSim. The paper reports metric gains for most configurations, a parameter increase below 1% for ViTs, and an ablation isolating the contributions of feature banks and sampling banks.

Significance. If the claimed gains hold, the bank mechanism is a cheap and generic architectural add-on to transformer-based dense-prediction decoders, with a negligible parameter overhead and a plausible mechanism: giving each decoding block global context from all feature scales. The paper has several concrete strengths: it tests multiple encoder families, reports footprint analysis in Table III, evaluates on synthetic ground truth for HyperSim, and includes a component ablation in Table IV. The central weakness is that the main natural-image evidence is obtained by training and evaluating against DepthAnythingV2 pseudo-labels rather than ground truth, so the headline claim of improved depth estimation on natural images is not yet substantiated; the mixed regressions for ViT-B and the absence of error bars further limit the strength of the claim.

major comments (4)
  1. [Section IV-A, Tables I] The paper states in Section IV-A: 'We train and evaluate on DepthAnythingV2 pseudo-labels as opposed to the ground truth in the evaluation datasets.' Because the same pseudo-labels supervise training, the No Bank versus Bank comparisons on NYUv2, Mannequin Challenge, and IRS in Table I measure the models' agreement with DepthAnythingV2's depth predictions, not their accuracy against real ground-truth depth. Consequently, the abstract's unqualified claim that banks 'improve performance on depth estimation' for natural images is not established by the current evidence. I request either an additional ground-truth evaluation on NYUv2 and, where available, the other natural-image benchmarks, or a carefully qualified claim that the gains are measured with respect to pseudo-label agreement.
  2. [Section IV-B, Table I, ViT-B on NYUv2] Table I shows that for ViT-B with DINOv2 weights, the AbsRel on NYUv2 worsens from 0.167 to 0.175 when banks are added, even though delta1 improves from 0.826 to 0.848. Section IV-B states that banks 'are able to cause increase in all metrics across the board, albeit with a few exceptions' and then mentions only the AbsRel deteriorations on Mannequin Challenge and IRS. The NYUv2 AbsRel regression is omitted from that discussion. Since AbsRel is a headline depth metric and this is a regression on one of the principal benchmarks, the omission is material to the central claim and should be explicitly reported and analyzed.
  3. [Section IV-B, Table II, ViT-B on HyperSim] In the ground-truth HyperSim evaluation, ViT-B shows a mixed result that is not fully disclosed in the text: AbsRel worsens from 0.134 to 0.138 and delta3 drops from 0.979 to 0.978, even though delta1 improves from 0.923 to 0.924. The text mentions the delta3 deterioration but not the AbsRel one. Given that this table is the only ground-truth natural-image-scale evidence, the complete set of regressions should be reported and the narrative of consistent gains for ViT-B should be reconciled with these numbers.
  4. [Tables I, II, IV] All experiments appear to be single runs with no error bars, seeds, or statistical significance assessment. Several of the reported differences are very small (for example, ViT-B HyperSim delta1 of 0.923 to 0.924, and delta3 of 0.979 to 0.978), so it is not possible to determine whether the bank gains are reliable or are within run-to-run noise. I request that the main comparisons be repeated with at least three seeds and reported as mean +/- standard deviation, or that some other significance evidence be provided.
minor comments (5)
  1. [Section III-B2, Eq. (6)] The notation 'GS ↑ / ↓ (Xin, Xref) : R → R' is not a well-formed function signature; Xin and Xref are tensors, not scalars, and the arrow notation is ambiguous. Please specify the input and output tensor shapes, or introduce the function as mapping between feature maps rather than between real numbers.
  2. [Section III, Table III] The bank channel widths used in Table III (64, 128, 256) are reported only in the table and are not connected to the method description. Please state how these widths are chosen for each encoder and whether they were tuned per architecture.
  3. [Section IV-A] The training procedure reports only the starting learning rate and the L1 loss. For reproducibility, the learning-rate schedule, batch size, input resolution, number of iterations for the large-scale mixed dataset, and the exact composition of the 1M-image training mix should be specified.
  4. [Figure 4] The qualitative comparison in Figure 4 would be clearer if the captions identified which encoder and checkpoint were used, and whether the 'Bank' results come from the model with the best metrics or from a model selected for qualitative behavior.
  5. [References] The paper cites the authors' own prior work [31] in the related work without explaining its connection to the proposed bank mechanism; a sentence clarifying what is taken from [31] and what is new in the present paper would improve the self-containedness of the presentation.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the bank architecture is an empirical, fit-free modification whose comparisons are self-contained; the only self-citation is non-load-bearing.

full rationale

This paper is an empirical architecture study with no derivation chain whose conclusion is assumed in its premises. The bank mechanism is defined constructively (Equations 2-7) as shared tensors generated from encoder features and consumed by decoder blocks, and performance is assessed by comparing trained models with and without banks on depth benchmarks. The only self-citation is reference [31], cited in Section I as an example of a convolutional decoder; it does not define banks, guided sampling, or any evaluation metric and is therefore not load-bearing. The use of DepthAnythingV2 pseudo-labels for training and evaluation on natural-image datasets (Section IV-A) is a benchmarking choice that may affect external validity, but it is applied identically to both the No Bank and Bank conditions, so the reported difference is an empirical ablation result rather than a quantity forced by construction. No equation in the paper defines a prediction in terms of the outcome it claims to predict, and no fitted parameter is relabeled as a prediction. Consequently, no circular step is identified.

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

The central claim rests on pretrained encoders, the pseudo-label proxy for ground truth, and hand-chosen bank widths. The two introduced architectural components (feature and sampling banks) are evidenced only by in-paper experiments, with no released code or external falsifiable handle.

free parameters (2)
  • Bank channel width = 64 (ViT-S), 128 (ViT-B), 256 (RepViT m1.1 and m2.3)
    Chosen by hand from Table III and not ablated; it determines the capacity and footprint of the shared bank tensors.
  • Starting learning rate = 5e-5
    Reported in Section IV-A; chosen by hand, not ablated, but standard for fine-tuning.
assumptions (3)
  • domain assumption Pretrained ViT and RepViT encoders provide suitable source features for dense depth prediction.
    All experiments start from DINOv2 or ImageNet-21k pretrained weights (Section IV-A); the bank mechanism is validated only on top of these upstream representations.
  • ad hoc to paper DepthAnythingV2 pseudo-labels are a valid proxy for ground truth in natural-image depth evaluation.
    Section IV-A explicitly evaluates NYUv2, Mannequin Challenge, and IRS against pseudo-labels rather than ground truth; the validity of this proxy is not independently checked.
  • domain assumption L1 loss alone is sufficient to reveal architectural differences between decoders.
    Section IV-A says depth-specific losses were beyond scope; relative comparisons may still hold, but absolute metric values are not optimized.
invented entities (2)
  • Feature bank B_feat
    purpose: Shared tensor generated from all encoder feature maps and multiplied into each decoder block's fused features to reweight them (Eq. 4-5).
    Its benefit is shown only by in-paper ablation; no external predictions or released artifacts allow independent verification.
  • Sampling bank B_sample
    purpose: Shared tensor used as guidance for dynamic up/downsampling inside each decoder block (Eq. 6-7), replacing bilinear upsampling.
    Built on DySample; the claimed qualitative and metric gains are only evidenced in-paper, with no released implementation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rethinking Encoder-Decoder Flow Through Shared Structures." pith.science (2026). https://pith.science/paper/RWASWMTI

@misc{pith2026250114535,
  author       = {Pith},
  title        = {Pith review of: Rethinking Encoder-Decoder Flow Through Shared Structures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RWASWMTI}},
  note         = {Machine review of arXiv:2501.14535}
}
read the original abstract

Dense prediction tasks have enjoyed a growing complexity of encoder architectures, decoders, however, have remained largely the same. They rely on individual blocks decoding intermediate feature maps sequentially. We introduce banks, shared structures that are used by each decoding block to provide additional context in the decoding process. These structures, through applying them via resampling and feature fusion, improve performance on depth estimation for state-of-the-art transformer-based architectures on natural and synthetic images whilst training on large-scale datasets.

Figures

Figures reproduced from arXiv: 2501.14535 by the authors.

Figure 1
Figure 1. Accuracy (δ1) vs parameters for architectures trained on large-scale dataset and tested on NYUv2. The introduction of our proposed banks can make a ViT-S based model almost match the performance of a ViT-B based model, whilst introducing a negligible amount of parameters. Research around encoders, specifically ViT-based ones, has either centered around introducing convolutional priors or altering the atten￾tion mech… view at source ↗
Figure 2
Figure 2. Encoder-decoder flow with banks (a) shown with RefineNet decoder block (b) and our block with banks (c). [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. DySample (a) reconfigured to support guided sampling via reference [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative results of RepViT m2 3 on HyperSim dataset. The usage of banks recovers artifacts and improves geometric accuracy of edges. in our block we downsample the blocks with the feature maps a guidance tensor. In a general formulation, we can define a guided sampl…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 26 canonical work pages

  1. [31]

    Mehmet Kerim Yucel, Valia Dimaridou, Anastasios Drosou, and Al- bert Saa-Garriga, ”Real-time monocular depth estimation with sparse supervision on mobile”, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2428–2437, 2021

    work page 2021

  2. [1]

    Vision trans- formers for dense prediction,

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun, “Vision trans- formers for dense prediction,” in Proceedings of the IEEE/CVF interna- tional conference on computer vision, 2021, pp. 12179–12188

    work page 2021

  3. [2]

    Dinov2: Learning robust visual features without supervision,

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al., “Dinov2: Learning robust visual features without supervision,” arXiv preprint arXiv:2304.07193, 2023

    arXiv 2023

  4. [3]

    RefineNet: Multi-path refinement networks for high-resolution semantic segmenta- tion,

    Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid, “RefineNet: Multi-path refinement networks for high-resolution semantic segmenta- tion,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1925–1934

    work page 2017

  5. [4]

    Eva: Exploring the limits of masked visual representation learning at scale,

    Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao, “Eva: Exploring the limits of masked visual representation learning at scale,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2023, pp. 19358–19369

    work page 2023

  6. [5]

    Depth anything: Unleashing the power of large- scale unlabeled data,

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao, “Depth anything: Unleashing the power of large- scale unlabeled data,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 10371–10381

    work page 2024

  7. [6]

    Repurposing diffusion-based image generators for monocular depth estimation,

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler, “Repurposing diffusion-based image generators for monocular depth estimation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 9492–9502

    work page 2024

  8. [7]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weis- senborn, Xiaohua Zhai, ThomasUnterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929, 2020

    arXiv 2010

Show all 32 references
  1. [8]

    Midasv3. 1–a model zoo for robust monocular relative depth estimation,

    Reiner Birkl, Diana Wofk, and Matthias M ¨uller, “Midasv3. 1–a model zoo for robust monocular relative depth estimation,” arXiv preprint arXiv:2307.14460, 2023

    2023 arXiv

  2. [9]

    Zoedepth: Zero-shot transfer by combining relative and metric depth,

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M ¨uller, “Zoedepth: Zero-shot transfer by combining relative and metric depth,” arXiv preprint arXiv:2302.12288, 2023

    2023 arXiv

  3. [10]

    Depthfm: Fast monocular depth estimation with flow matching,

    Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Bj ¨orn Ommer, “Depthfm: Fast monocular depth estimation with flow matching, ”arXiv preprint arXiv:2403.13788, 2024

    2024 arXiv

  4. [11]

    Bins- former: Revisiting adaptive bins for monocular depth estimation,

    Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang, “Bins- former: Revisiting adaptive bins for monocular depth estimation,”IEEE Transactions on Image Processing, 2024

    2024

  5. [12]

    Ad- abins: Depth estimation using adaptive bins,

    Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka, “Ad- abins: Depth estimation using adaptive bins,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp.4009–4018

    2021

  6. [13]

    Shvit: Single-head vision transformer with memory efficient macro design,

    Seokju Yun and Youngmin Ro, “Shvit: Single-head vision transformer with memory efficient macro design,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 5756–5767

    2024

  7. [14]

    Fast vision transformers with hilo attention,

    Zizheng Pan, Jianfei Cai, and Bohan Zhuang, “Fast vision transformers with hilo attention,” Advances in Neural Information Processing Sys- tems, vol. 35, pp. 14541–14554, 2022

    2022

  8. [15]

    Swin transformer: Hier- archical vision transformer using shifted windows,

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo, “Swin transformer: Hier- archical vision transformer using shifted windows,” in Proceedings of the IEEE/CVFinternational conference on computer vision, 2021, pp.10012–10022

    2021

  9. [16]

    Swin-unet: Unet-like pure transformer for medical image segmentation,

    Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang, “Swin-unet: Unet-like pure transformer for medical image segmentation,” in European conference on computer vision. Springer, 2022, pp. 205–218

    2022

  10. [17]

    Efficientformer: Vision trans- formers at mobilenet speed,

    Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren, “Efficientformer: Vision trans- formers at mobilenet speed,” Advances in Neural Information Processing Systems, vol. 35, pp. 12934–12949, 2022

    2022

  11. [18]

    Swiftformer: Efficient additive attention for transformer-based real-time mobile vision applications,

    Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fa-had Shahbaz Khan, “Swiftformer: Efficient additive attention for transformer-based real-time mobile vision applications,” in Proceedings of the IEEE/CVF International Conference on Compute...

    2023

  12. [19]

    Fastvit: A fast hybrid vision transformer using struc- tural reparameterization,

    Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan, “Fastvit: A fast hybrid vision transformer using struc- tural reparameterization,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 5785–5795

    2023

  13. [20]

    Repvit: Revisiting mobile cnn from vit perspective,

    Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding, “Repvit: Revisiting mobile cnn from vit perspective,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15909–15920

    2024

  14. [21]

    Segformer: Simple and efficient design for se- mantic segmentation with transformers,

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandku-mar, Jose M Alvarez, and Ping Luo, “Segformer: Simple and efficient design for se- mantic segmentation with transformers,” Advances in neural information processing systems, vol. 34, pp. 12077–12090, 2021

    2021

  15. [22]

    Segmenter: Transformer for semantic segmentation,

    Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid, “Segmenter: Transformer for semantic segmentation,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp.7262–7272

    2021

  16. [23]

    More than encoder: Introducing transformer decoder to upsample,

    Yijiang Li, Wentian Cai, Ying Gao, Chengming Li, and Xiping Hu, “More than encoder: Introducing transformer decoder to upsample,” in 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2022, pp. 1597–1602

    2022

  17. [24]

    Denoising vision transformers,

    Jiawei Yang, Katie Z Luo, Jiefeng Li, Kilian Q Weinberger, Yonglong Tian, and Yue Wang, “Denoising vision transformers,” arXiv preprint arXiv:2401.02957, 2024

    2024 arXiv

  18. [25]

    Learning to upsample by learning to sample,

    Wenze Liu, Hao Lu, Hongtao Fu, and Zhiguo Cao,“Learning to upsample by learning to sample,” in Proc. IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    2023

  19. [26]

    Imagenet-21k pretraining for the masses,

    Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor, “Imagenet-21k pretraining for the masses,” 2021

    2021

  20. [27]

    Indoor segmentation and support inference from rgbd images,

    Pushmeet Kohli, Nathan Silberman, Derek Hoiem and Rob Fergus, “Indoor segmentation and support inference from rgbd images,” in ECCV , 2012

    2012

  21. [28]

    Learning the depths of moving people by watching frozen people,

    Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman, “Learning the depths of moving people by watching frozen people,” in Proc. Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  22. [29]

    Irs: A large naturalistic indoor robotics stereo dataset to train deep models for disparity and surface normal estimation,

    Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu, “Irs: A large naturalistic indoor robotics stereo dataset to train deep models for disparity and surface normal estimation,”in 2021 IEEE International Conference on Multimedia and Expo (ICME). IE...

    2021

  23. [30]

    Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding,

    Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind, “Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding,” in International Conference on Computer Vision (ICCV) 2021, 2021

    2021

  24. [32]

    The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale,

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al., “The open images dataset v4: Unified image classification, object detection, and visual relationship detection...

    1956

Pith tools

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