Pith. sign in

REVIEW 3 major objections 5 minor 45 references

DPNeXt: A Lightweight Multi-Scale Feature Fusion Framework for Efficient ViT-Based Multi-Task Dense Prediction

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read DPNeXt, a lightweight decoder with depthwise separable inverted fusion, replaces the standard DPT to deliver top multi-task dense prediction results on Cityscapes and NYUv2 with 78.6% fewer trainable parameters.

desk verdict Solid architecture paper whose efficiency claims hold up internally, but the SOTA comparison on Cityscapes rests on unverifiable baseline numbers and a thin margin. read the letter →

arxiv 2607.16012 v1 pith:PARVOTX3 submitted 2026-07-17 cs.CV cs.AIcs.RO

Jehun Kang , Jungha Wang , Youngjun Hwang , David Hyunchul Shim This is my paper
classification cs.CVcs.AIcs.RO
keywords multi-tasklearningdensepredictionTransformerDINOv2featurefusiondepthwiseseparableconvolutionboundaryguidanceefficientinference
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

The paper proposes DPNeXt, a lightweight decoder that replaces the standard Dense Prediction Transformer (DPT) in multi-task dense prediction. Its central claim is that a frozen DINOv2-Reg backbone plus DPNeXt's fusion blocks and a training-only boundary guidance strategy yields better semantic segmentation and depth estimation than prior multi-task models while using far fewer trainable parameters. DPNeXt-S cuts trainable parameters by 78.6% versus standard DPT and runs fastest among compared models on a laptop GPU, according to the paper's measurements. If the results hold, efficient ViT-based multi-task perception becomes practical for resource-constrained robotics.

What carries the argument

The load-bearing mechanism is the DDSIF block: a dual depthwise separable inverted bottleneck that replaces the dense-convolution residual fusion units of DPT, performing two depthwise spatial filters separated by a pointwise expansion (ratio r=2) to fuse multi-scale features with minimal parameters. Working alongside it is the Isotropic Projection Adapter (IPA), which avoids the artificial channel expansion used by CNN-style reassembly by projecting all encoder layers into one 256-dimension isotropic fusion space. The Multi-Task Boundary Guidance (MTBG) strategy supplies the task-regularization mechanism: a training-only auxiliary boundary head and boundary-masked segmentation/depth losses

What would settle it

Use official checkpoints of M2H, SwinMTL, and other baselines, and re-evaluate them on Cityscapes using the same 518×518 training crops, sliding-window 518×518 stitching at 1024×2048, CREStereo depth range [0.001, 80], and the same JPS formula; if any baseline surpasses JPS 0.867 (DPNeXt-B) or 0.858 (DPNeXt-S), the state-of-the-art claim fails.

Watch

Extended reading notes

Core claim

DPNeXt is a streamlined multi-scale feature fusion decoder proposed as an efficient alternative to the standard Dense Prediction Transformer. Its Isotropic Projection Adapter projects four layers of a frozen DINOv2-Reg backbone into a shared 256-dimension fusion space without CNN-style channel widening; its Dual Depthwise Separable Inverted Fusion blocks aggregate the pyramidal features using two depthwise separable stages with expansion ratio 2. Multi-Task Boundary Guidance adds a training-only boundary head and boundary-restricted losses, with masks from labels or Canny-derived boundaries. The paper reports DPNeXt-S outperforms prior SOTA MTL models on Cityscapes, reduces trainable paramet

Load-bearing premise

The 'best among compared methods' claim depends on the comparison baselines being evaluated under fair, equivalent protocols; some prior results come from original papers, third-party reproductions, or checkpoints the authors found underperform, so an apples-to-apples re-run could change the ranking.

Editorial extensions

If this is right

  • With a frozen DINOv2-Reg encoder, the DPNeXt decoder produces a shared fused feature that supports multiple dense tasks at a fraction of DPT's trainable parameter cost.
  • MTBG offers a recipe for injecting geometric structure into multi-task decoders without adding runtime overhead, since the auxiliary boundary head is discarded at inference.
  • The measured inference-speed advantage over higher-FLOP baselines suggests that pure-CNN decoders are better aligned with actual hardware execution than custom attention or scan operations.
  • If the reported benchmark numbers hold, ViT-based multi-task dense prediction becomes viable on laptop-class hardware, which matters for on-robot perception.

Reading between the lines

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

  • Because the IPA alone already cuts parameters substantially while preserving performance, the channel-widening step in standard DPT-style reassembly may be largely redundant for frozen isotropic VFMs; this is worth testing on other isotropic backbones.
  • MTBG's boundary masks are derived from segmentation labels via Canny plus dilation; if the performance gain comes mainly from hard-example emphasis rather than geometric consistency, the strategy should transfer to other tasks such as monocular depth in video.
  • The flipped efficiency trade-off (higher GFLOPs yet faster wall-clock speed) implies that FLOPS-based comparisons in MTL papers can mislead; a community-wide adoption of target-hardware latency reporting would change how efficiency claims are judged.
  • Applying DPNeXt to streaming or open-vocabulary perception, as the paper suggests for future work, would test whether the compact fused feature retains semantic and geometric information beyond the two supervised tasks.
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

3 major / 5 minor

Summary. This manuscript proposes DPNeXt, a lightweight decoder for multi-task dense prediction (semantic segmentation, monocular depth estimation) built on top of a frozen DINOv2-Reg ViT encoder. The decoder replaces DPT's reassembly blocks with an Isotropic Projection Adapter (IPA) and Dual Depthwise Separable Inverted Fusion (DDSIF) blocks, and introduces a Multi-Task Boundary Guidance (MTBG) training strategy that applies boundary-aware losses to segmentation and depth without adding inference-time branches. Experiments on Cityscapes and NYUv2 report that DPNeXt-S and DPNeXt-B achieve the best JPS among the compared models, with DPNeXt-S reducing trainable parameters by 78.6% relative to a DPT-DINOv2-S baseline and reaching 51.0 FPS on a laptop GPU. Ablations show that each proposed component improves the aggregate metric.

Significance. If the reported comparisons are validated, the contribution is practically significant: it demonstrates that a frozen VFM backbone plus a carefully designed, hardware-friendly convolutional decoder can match or beat much heavier MTL decoders, with a large reduction in trainable parameters and no inference-time penalty from the boundary guidance. The design rationale is clearly motivated by the isotropic nature of DINOv2 features, and the ablation table gives a coherent decomposition of the gains. The MTBG strategy is simple, annotation-free (label-derived boundaries), and shown to improve the JPS, which makes it a useful ingredient for future MTL systems. The main caveat is that the SOTA claim is currently entangled with the comparability of baseline numbers, especially M2H, and with the small margins between methods; hence the practical significance is contingent on resolving this.

major comments (3)
  1. [Sec. IV-A, Tables I-II] The central claim that DPNeXt-S outperforms prior SOTA on Cityscapes rests on Table I, where the strongest prior model M2H is marked † (code/checkpoints unavailable) and its JPS (0.850) is only 0.008 below DPNeXt-S (0.858). In Table II, M2H's official checkpoints produce 'lower than reported' values (‡), indicating that published M2H numbers can be optimistic. The Cityscapes baseline values are assembled from 'original published results, third-party reproductions, and our own empirical evaluations' with no unified protocol for input resolution, depth label source, or inference post-processing (sliding-window/Gaussian blending is used only for DPNeXt and DPT-DINOv2-S). Unless M2H is evaluated under the same protocol, the 'outperforms prior SOTA' claim is not established. This is load-bearing and needs correction, either by providing an apples-to-apples rerun of M2H on Cityscapes or by tem
  2. [Sec. IV-A, Table I] No error bars or multiple-seed results are reported. Margins over the closest competitors are small: e.g., DPNeXt-S's JPS advantage over M2H is 0.008 on Cityscapes, and DPNeXt-B's JPS advantage over M2H on NYUv2 is 0.002 (Table II). Given the sensitivity of dense-prediction metrics to random seed, learning-rate schedule, and crop sampling, these differences may be within run-to-run variation. Please report mean and standard deviation over at least three independent runs for the main comparisons, or justify the single-run protocol.
  3. [Sec. IV-A, Eq. (14)] The 'best overall performance' rankings are based on the author-defined JPS, an equal-weighted combination of normalized mIoU and normalized RMSE. This aggregate is not a standard benchmark metric and can change rankings depending on the chosen normalization and task weighting. For example, on NYUv2, DPNeXt-B improves mIoU by 0.47 pp over M2H but reduces RMSE by 0.0028 m; whether this constitutes 'best' depends on the application's task preference. The paper should present a sensitivity analysis or at least explicitly discuss the arbitrariness of the aggregate and show that conclusions are robust to reasonable alternative weightings.
minor comments (5)
  1. [Table III] The M2H-Small row has a typesetting issue: the GFLOPs and FPS values appear run together (e.g., '51.620.72'). Please format as separate columns.
  2. [Sec. III-C] When boundary masks are derived from semantic labels with Canny edge detection and dilation, the exact kernel size, dilation radius, and threshold are not given. Providing these values would improve reproducibility.
  3. [Sec. IV-E] The ablation results in Table IV are trained for 160 epochs at batch size 32, whereas the main models use 400 epochs and batch size 8. The comparisons within the table are valid, but the text should explicitly warn readers not to compare the Table IV accuracy numbers against Table I.
  4. [Sec. III-C.2] The boundary-aware losses L_bas and L_bad reuse L_seg and L_depth, but it is not stated how the OHEM selection (Eq. 10) and SiLog normalization (Eq. 12) interact with the restriction to M_bound. Please clarify whether hard-pixel mining is performed over boundary pixels only or globally before masking.
  5. [Table I] M2H's parameter count is listed as '-' in Table I but as 81.54M in Table III. Please reconcile these entries.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; all accuracy claims rest on external benchmarks and no fitted quantity is presented as a prediction.

full rationale

The paper's derivation chain is self-contained and empirical. DPNeXt's decoder (IPA + DDSIF) is a proposed architecture whose components are defined by their own equations (Eqs. 1-5) and compared against standard DPT [9] and other external baselines; no component is defined in terms of a target result. The MTBG losses (Eqs. 6-8) are training objectives, and the total loss (Eq. 13) is a declared linear combination with reported weights, not a fitted quantity later relabeled as a prediction. The JPS metric (Eq. 14) is explicitly introduced as the authors' own summary measure, so using it to rank methods is transparent rather than circular. There are no self-citations by the present authors in the reference list, and the frozen DINOv2-Reg backbone is an external, independently published model. The paper does disclose a comparability limitation in Sec. IV-A and Tables I-II: some baseline numbers come from original papers, third-party reproductions, or reimplementations, and for M2H official checkpoints are unavailable (Cityscapes) or produce lower values than reported (NYUv2). This is a threat to the fairness of the SOTA comparison and to claim strength, but it is a benchmarking/correctness concern, not a circular derivation: the claimed predictions are still measured against held-out benchmark labels and are not forced by construction. No reduction of any result to its inputs was found.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. The contribution is architectural, so the ledger is dominated by hand-chosen hyperparameters (loss weights, OHEM threshold, fusion dimension, expansion ratio) and domain assumptions about benchmark validity and baseline comparability.

free parameters (4)
  • Loss weights λ_seg, λ_depth, λ_bound, λ_bas, λ_bad, λ_normal = Cityscapes: (1.0, 1.0, 50.0); NYUv2-S: (1.5, 1.0, 50.0, 10.0); NYUv2-B: (1.0, 2.0, 50.0, 10.0); λ_bas=λ_bad=0.4
    Tuned separately per dataset and model scale; these directly define the total objective (Eq. 13) and differ between DPNeXt-S and DPNeXt-B on NYUv2.
  • OHEM threshold τ and min kept-pixel ratio = τ = 0.9, min ratio = 1/8
    Hard-pixel selection threshold in the segmentation loss (Eq. 9-10); affects which pixels contribute to training.
  • Fusion dimension d_fusion and DDSIF expansion ratio r = d_fusion = 256, r = 2
    Architectural constants chosen by hand; not fitted to target metrics but control parameter count and capacity.
  • Depth normalization range and λ_silog = d_min=0.001, d_max=80.0 (Cityscapes) / 10.0 (NYUv2); λ_silog=0.5
    Depth clipping and log-normalization (Eq. 12) and SiLog variance term (Eq. 11) are standard choices but manually set.
assumptions (5)
  • domain assumption Frozen DINOv2-Reg backbone features are sufficient for both semantic segmentation and depth estimation.
    The decoder does not fine-tune the backbone; if frozen VFM features lack task-specific geometric detail, the architecture's ceiling is set by the backbone (Sec. III-A.1).
  • domain assumption Cityscapes depth maps produced by CREStereo are valid pseudo-ground truth for evaluating monocular depth.
    Depth RMSE is computed against stereo-estimated depth, not LiDAR; systematic errors in CREStereo would propagate to all depth numbers (Sec. IV-A).
  • domain assumption Canny edge detection followed by dilation on semantic labels produces boundary masks suitable for geometric supervision.
    MTBG derives boundary labels from segmentation masks, following the PIDNet protocol; noisy masks would encode noise into the auxiliary losses (Sec. III-C).
  • ad hoc to paper The JPS metric (equal-weight combination of normalized mIoU and depth RMSE) is a valid aggregate ranking criterion.
    JPS is defined in Eq. 14 by the authors; "best overall performance" claims depend on accepting this weighting.
  • domain assumption Reported baseline numbers from prior papers/reproductions are accurate and comparable to the authors' protocol.
    Tables I-II mix original published results, third-party reproductions, and the authors' own baselines; footnotes note unavailable checkpoints and lower official M2H results (Sec. IV-A).

how reviews work

0 comments
Cite this review

Pith. "Pith review of DPNeXt: A Lightweight Multi-Scale Feature Fusion Framework for Efficient ViT-Based Multi-Task Dense Prediction." pith.science (2026). https://pith.science/paper/PARVOTX3

@misc{pith2026260716012,
  author       = {Pith},
  title        = {Pith review of: DPNeXt: A Lightweight Multi-Scale Feature Fusion Framework for Efficient ViT-Based Multi-Task Dense Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PARVOTX3}},
  note         = {Machine review of arXiv:2607.16012}
}
read the original abstract

Multi-Task Learning (MTL) in robotics perception systems supports comprehensive 3D spatial scene understanding by integrating semantic segmentation and depth estimation. While Vision Foundation Models (VFMs) are increasingly adopted as robust feature encoders, existing decoding strategies present a critical bottleneck. To address this, we propose DPNeXt, a streamlined multi-scale feature fusion decoder and efficient alternative to the standard Dense Prediction Transformer (DPT). DPNeXt uses dual depthwise separable inverted bottlenecks to improve frozen VFM utilization through fusion-centric decoding and independent task modularization. To further mitigate negative inductive transfer between tasks, we introduce the Multi-Task Boundary Guidance (MTBG) strategy. Unlike prior boundary-aware methods that add fusion modules or gating, MTBG applies symmetric boundary-focused supervision to encourage geometric consistency without extra annotation or inference cost. Experiments on Cityscapes show that DPNeXt-S outperforms prior state-of-the-art (SOTA) MTL models, while DPNeXt-B further improves the overall performance and achieves the best results among the compared methods. On NYUv2, DPNeXt-B also achieves the best semantic segmentation and depth estimation results among the compared methods while requiring substantially fewer trainable parameters than prior large-scale MTL models. Compared with the standard DPT, DPNeXt-S reduces trainable parameters by 78.6% and achieves the fastest inference speed among the compared models on resource-constrained laptop hardware. The source code, model checkpoints, and a demo video will be made available at https://github.com/kangjehun/DPNeXt.

Figures

Figures reproduced from arXiv: 2607.16012 by the authors.

Figure 1
Figure 1. Overview of the proposed MTL framework. The architec￾ture operates in two main stages. In Stage I, a frozen VFM backbone (Blue) extracts rich representations, which the DPNeXt decoder (Orange) aggregates into a shared fused feature (Gray) through multi-scale fusion. This feature aggregates complementary multi-scale cues and can serve as a compact latent representation for downstream perception modules. In Stage II, … view at source ↗
Figure 2
Figure 2. Detailed Architecture of the Proposed MTL Framework. (a) The frozen DINOv2-Reg backbone (Blue) extracts robust isotropic feature tokens from the input image. (b) The proposed DPNeXt decoder (Orange) aggregates these features using the Isotropic Projection Adapter (IPA) for dimension alignment and resolution resizing. Subsequently, the Dual Depthwise Separable Inverted Fusion (DDSIF) block efficiently fuses the resul… view at source ↗
Figure 3
Figure 3. Visualization of DINOv2 Attention Maps. We visualize the attention response of the [CLS] token across Layers 3, 6, 9, and 12. Register tokens were removed for spatial readability. (a) Layer 3 shows sparse and localized attention patterns. (b) Layer 6 begins to highlight object-level regions. (c) Layer 9 covers broader contextual areas. (d) Layer 12 exhibits more global scene-level responses. where the subscript sk ∈… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Structure of DPNeXt Decoder Components. Left: We contrast the widely adopted artificial channel expansion strategy (Green) with the original DPT approach (Orange, Left) used in our IPA block. By avoiding redundant channel widening with pointwise convolution (PW), IPA p…
Figure 5
Figure 5. Figure 5: Qualitative Results on the Cityscapes Dataset. We compare the predictions of SwinMTL, DPT-DINOv2-S, and our DPNeXt-S against the GT across diverse urban scenes. For enhanced visualization, specific regions are cropped and magnified. The original crop regions are indica…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 4 linked inside Pith

  1. [1]

    Bev- former: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers,

    Z. Li, W. Wang, H. Li, E. Xie, C. Sima, T. Lu, Q. Yu, and J. Dai, “Bev- former: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 3, pp. 2020–2036, 2024

    2020

  2. [2]

    Vadv2: End-to-end vectorized autonomous driving via probabilistic planning,

    S. Chen, B. Jiang, H. Gao, B. Liao, Q. Xu, Q. Zhang, C. Huang, W. Liu, and X. Wang, “Vadv2: End-to-end vectorized autonomous driving via probabilistic planning,”arXiv preprint arXiv:2402.13243, 2024

    arXiv 2024

  3. [3]

    Scene as occupancy,

    W. Tong, C. Sima, T. Wang, L. Chen, S. Wu, H. Deng, Y . Gu, L. Lu, P. Luo, D. Linet al., “Scene as occupancy,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 8406–8415

    2023

  4. [4]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafsonet al., “Sam 2: Segment anything in images and videos,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 28 085–28 128

    2025

  5. [5]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,”Advances in Neural Information Processing Systems, vol. 37, pp. 21 875–21 911, 2024

    2024

  6. [6]

    Multi-task learning for dense prediction tasks: A survey,

    S. Vandenhende, S. Georgoulis, W. Van Gansbeke, M. Proesmans, D. Dai, and L. Van Gool, “Multi-task learning for dense prediction tasks: A survey,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 7, pp. 3614–3633, 2021

    2021

  7. [7]

    Multitask learning,

    R. Caruana, “Multitask learning,”Machine learning, vol. 28, no. 1, pp. 41–75, 1997

    1997

  8. [8]

    Cross-stitch net- works for multi-task learning,

    I. Misra, A. Shrivastava, A. Gupta, and M. Hebert, “Cross-stitch net- works for multi-task learning,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3994–4003

    2016

Show all 45 references
  1. [9]

    Vision transformers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 12 179–12 188

    2021

  2. [10]

    Invpt++: Inverted pyramid multi-task transformer for visual scene understanding,

    H. Ye and D. Xu, “Invpt++: Inverted pyramid multi-task transformer for visual scene understanding,”IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 12, pp. 7493–7508, 2024

    2024

  3. [11]

    Taskprompter: Spatial-channel multi-task prompting for dense scene understanding,

    ——, “Taskprompter: Spatial-channel multi-task prompting for dense scene understanding,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  4. [12]

    Mtmamba: Enhancing multi-task dense scene understanding by mamba-based decoders,

    B. Lin, W. Jiang, P. Chen, Y . Zhang, S. Liu, and Y .-C. Chen, “Mtmamba: Enhancing multi-task dense scene understanding by mamba-based decoders,” inEuropean conference on computer vision. Springer, 2024, pp. 314–330

    2024

  5. [13]

    Multi-task dense prediction via mixture of low-rank experts,

    Y . Yang, P.-T. Jiang, Q. Hou, H. Zhang, J. Chen, and B. Li, “Multi-task dense prediction via mixture of low-rank experts,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 27 927–27 937

    2024

  6. [14]

    Swinmtl: A shared architecture for simultaneous depth estimation and semantic segmentation from monocular camera images,

    P. Taghavi, R. Langari, and G. Pandey, “Swinmtl: A shared architecture for simultaneous depth estimation and semantic segmentation from monocular camera images,” in2024 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 4957–4964

    2024

  7. [15]

    M2h: Multi-task learning with efficient window-based cross-task attention for monocular spatial perception,

    U. Udugama, G. V osselman, and F. Nex, “M2h: Multi-task learning with efficient window-based cross-task attention for monocular spatial perception,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 8067–8072

    2025

  8. [16]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  9. [17]

    Mobilenets: Efficient convolutional neural networks for mobile vision applications,

    A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam, “Mobilenets: Efficient convolutional neural networks for mobile vision applications,”arXiv preprint arXiv:1704.04861, 2017

    2017 arXiv

  10. [18]

    Mo- bilenetv2: Inverted residuals and linear bottlenecks,

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen, “Mo- bilenetv2: Inverted residuals and linear bottlenecks,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4510–4520

    2018

  11. [19]

    Efficientnet: Rethinking model scaling for con- volutional neural networks,

    M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for con- volutional neural networks,” inInternational conference on machine learning. PMLR, 2019, pp. 6105–6114

    2019

  12. [20]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11 976–11 986

    2022

  13. [21]

    Refinenet: Multi-path refinement networks for high-resolution semantic segmentation,

    G. Lin, A. Milan, C. Shen, and I. Reid, “Refinenet: Multi-path refinement networks for high-resolution semantic segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1925–1934

    2017

  14. [22]

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

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

    2010 arXiv

  15. [23]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022

    2021

  16. [24]

    Emerging properties in self-supervised vision trans- formers,

    M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision trans- formers,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9650–9660

    2021

  17. [25]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. Haziza, F. Massa, A. El-Noubyet al., “Dinov2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  18. [26]

    Vision transform- ers need registers,

    T. Darcet, M. Oquab, J. Mairal, and P. Bojanowski, “Vision transform- ers need registers,” inInternational conference on learning represen- tations, vol. 2024, 2024, pp. 2632–2652

    2024

  19. [27]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Confer- ence on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241

    2015

  20. [28]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2117–2125

    2017

  21. [29]

    Eff-unet: A novel architecture for semantic segmentation in unstructured environment,

    B. Baheti, S. Innani, S. Gajre, and S. Talbar, “Eff-unet: A novel architecture for semantic segmentation in unstructured environment,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 358–359

    2020

  22. [30]

    Swin transformer v2: Scaling up capacity and resolution,

    Z. Liu, H. Hu, Y . Lin, Z. Yao, Z. Xie, Y . Wei, J. Ning, Y . Cao, Z. Zhang, L. Donget al., “Swin transformer v2: Scaling up capacity and resolution,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 12 009–12 019

    2022

  23. [31]

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

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. 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. 10 371–10 381

    2024

  24. [32]

    Dino- foresight: Looking into the future with dino,

    E. Karypidis, I. Kakogeorgiou, S. Gidaris, and N. Komodakis, “Dino- foresight: Looking into the future with dino,”Advances in Neural Information Processing Systems, vol. 38, pp. 163 779–163 811, 2025

    2025

  25. [33]

    Pidnet: A real-time semantic segmentation network inspired by pid controllers,

    J. Xu, Z. Xiong, and S. P. Bhattacharyya, “Pidnet: A real-time semantic segmentation network inspired by pid controllers,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 19 529–19 539

    2023

  26. [34]

    Boundary-aware multitask learning for remote sensing imagery,

    Y . Wang, W. Ding, R. Zhang, and H. Li, “Boundary-aware multitask learning for remote sensing imagery,”IEEE Journal of selected topics in applied earth observations and remote sensing, vol. 14, pp. 951– 963, 2020

    2020

  27. [35]

    Training region-based object detectors with online hard example mining,

    A. Shrivastava, A. Gupta, and R. Girshick, “Training region-based object detectors with online hard example mining,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 761–769

    2016

  28. [36]

    Depth map prediction from a single image using a multi-scale deep network,

    D. Eigen, C. Puhrsch, and R. Fergus, “Depth map prediction from a single image using a multi-scale deep network,”Advances in neural information processing systems, vol. 27, 2014

    2014

  29. [37]

    Multi-task learning using uncertainty to weigh losses for scene geometry and semantics,

    A. Kendall, Y . Gal, and R. Cipolla, “Multi-task learning using uncertainty to weigh losses for scene geometry and semantics,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7482–7491

    2018

  30. [38]

    Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth es- timation and scene parsing,

    D. Xu, W. Ouyang, X. Wang, and N. Sebe, “Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth es- timation and scene parsing,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 675–684

    2018

  31. [39]

    Cross-task attention mecha- nism for dense multi-task learning,

    I. Lopes, T.-H. Vu, and R. De Charette, “Cross-task attention mecha- nism for dense multi-task learning,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 2329–2338

    2023

  32. [40]

    Three ways to improve semantic segmentation with self-supervised depth estimation,

    L. Hoyer, D. Dai, Y . Chen, A. Koring, S. Saha, and L. Van Gool, “Three ways to improve semantic segmentation with self-supervised depth estimation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 11 130–11 140

    2021

  33. [41]

    Inverted pyramid multi-task transformer for dense scene understanding,

    H. Ye and D. Xu, “Inverted pyramid multi-task transformer for dense scene understanding,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 514–530

    2022

  34. [42]

    Mtmamba++: Enhancing multi-task dense scene understanding via mamba-based decoders,

    B. Lin, W. Jiang, P. Chen, S. Liu, and Y .-C. Chen, “Mtmamba++: Enhancing multi-task dense scene understanding via mamba-based decoders,”IEEE Transactions on Pattern Analysis and Machine In- telligence, 2025

    2025

  35. [43]

    The cityscapes dataset for semantic urban scene understanding,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Be- nenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 3213–3223

    2016

  36. [44]

    Practical stereo matching via cascaded recurrent network with adaptive correlation,

    J. Li, P. Wang, P. Xiong, T. Cai, Z. Yan, L. Yang, J. Liu, H. Fan, and S. Liu, “Practical stereo matching via cascaded recurrent network with adaptive correlation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 16 263–16 272

    2022

  37. [45]

    Indoor segmenta- tion and support inference from rgbd images,

    N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, “Indoor segmenta- tion and support inference from rgbd images,” inEuropean conference on computer vision. Springer, 2012, pp. 746–760

    2012

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.