Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

VisLanding: Monocular 3D Perception for UAV Safe Landing via Depth-Normal Synergy

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read VisLanding claims that safe landing zones for UAVs are best found from predicted 3D geometry rather than from image texture, and that this choice buys large gains in cross-domain robustness.

desk verdict Useful SLZ adaptation with honest limitations, but the depth-normal synergy claim is untested; the cross-domain gain is likely confounded with pretraining. read the letter →

arxiv 2506.14525 v1 pith:CIP7IKJU submitted 2025-06-17 cs.CV cs.RO

classification cs.CVcs.RO
keywords UAVsafelandingzoneestimationmonoculardepthsurfacenormalcross-domaingeneralizationsemanticsegmentationdepth-normalsynergyMetric3DV2
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

VisLanding asks whether a UAV can find a safe place to land from a single RGB image by leaning on predicted 3D geometry rather than on appearance. The paper's claim is that adding a safe-landing-zone segmentation branch to the Metric3D V2 depth-and-normal model, and feeding the updated depth and normal maps into that branch during iterative refinement, yields a binary safe/unsafe segmentation that degrades far less than RGB-only segmenters when moved from wild terrain to urban scenes. On the WildUAV test split the method reaches 76.61% mIoU, and on the Semantic Drone cross-domain set it reaches 52.44% mIoU, compared with 39.61% for SafeUAV and 31.99% for SegFormer. The same predicted depth and normals are used to estimate the physical area of candidate landing zones, which matters for deciding whether a spot is big enough. The sympathetic reading is that geometry, not texture, is the transferable signal for landing safety.

What carries the argument

The load-bearing mechanism is the dual-flow iterative refinement inside the DPT module: the original depth-normal flow (three ConvGRU blocks) and a new SLZ flow (one ConvGRU) that concatenates the current depth, normal, and SLZ logits at each of $T$ steps. The geometric fusion forces the segmentation to be consistent with metric depth and surface orientation rather than with RGB texture. Two named supports make it work: Metric3D V2's canonical camera space, which removes scale ambiguity so depth is metric and transferable, and the depth-normal consistency loss, which keeps the two geometric branches coherent; the SLZ branch is trained with a temporally weighted cross-entropy loss that weights the safe class double.

What would settle it

On the Semantic Drone cross-domain set, ablate the SLZ flow's geometric input: if a version that receives only RGB features (no depth or normal conditioning) closes most of the 12-20 point mIoU gap over SafeUAV and SegFormer, then the claimed depth-normal synergy, rather than the dual-flow architecture or training scheme, is what drives the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that monocular safe-landing-zone estimation is best treated as a binary segmentation task supervised jointly with metric depth and surface normals, all refined in a shared recurrent optimization loop. VisLanding keeps Metric3D V2's canonical-space depth-normal flow intact and adds a lightweight SLZ flow whose ConvGRU receives the updated depth and normal maps at every iteration, so the segmentation is repeatedly re-grounded in geometry. The result is a model that trades a few points of in-domain mIoU against a large gain in cross-domain robustness: a drop of roughly 25-33% in key metrics versus 42-64% for the RGB-only baselines. The paper also derives a closed-form per-pixel area formula, $A_{\text{actual}} = D^2 / (f_x f_y |n_z|)$, that corrects for surface tilt using the predicted normal, and demonstrates on real flight data that area estimates improve as the drone approaches the target.

Load-bearing premise

The whole cross-domain advantage rests on the assumption that Metric3D V2's zero-shot depth and normal predictions stay accurate for UAV viewpoints and for unseen urban scenes, since the paper never measures that accuracy on the Semantic Drone test set.

Editorial extensions

If this is right

  • A UAV with only a camera can get metric landing-zone area estimates, not just a safe/unsafe mask, using the depth and normals it already predicts.
  • Cross-domain robustness means a model tuned on natural terrain can be deployed in urban settings without retraining on each new scene.
  • The dual-flow design keeps the zero-shot depth-normal foundation intact, so the SLZ branch inherits its generalization instead of fighting it.
  • Even a single optimization step ($T=1$) retains most of the cross-domain gain, giving a real-time/accuracy trade-off for onboard compute.

Reading between the lines

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

  • The same conditioning trick—feeding updated geometry into a segmentation head—could transfer to other safety tasks like ground traversability or obstacle-free path planning, provided a depth-normal backbone is available.
  • The area formula assumes locally planar ground and an accurate normal z-component; on rough or sloped terrain the estimates will be biased, so a planarity check on predicted normals could be added before trusting the area.
  • Because the in-domain numbers are lower than RGB baselines, the right use of this model is deployment under domain shift; a test on more than one cross-domain target would show whether the advantage is general or specific to Semantic Drone.
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

5 major / 6 minor

Summary. The paper presents VisLanding, a framework that extends Metric3D V2 with a safe landing zone (SLZ) segmentation branch, using depth and normal predictions from the DPT iterative optimization as additional inputs to a ConvGRU-based SLZ flow. The authors annotate WildUAV map images for SLZ, fine-tune the model, and evaluate it on held-out WildUAV images (in-domain) and on Semantic Drone images with author-constructed SLZ labels (cross-domain). Table I reports 76.61% mIoU in-domain and 52.44% mIoU cross-domain, with the latter outperforming SafeUAV and SegFormer baselines. The paper also estimates landing-zone areas via depth and normal projection. The main claimed contribution is that "depth-normal synergy" drives cross-domain generalization, but this mechanism is not isolated experimentally.

Significance. If the generalization mechanism were supported, this would be a practical and useful result: replacing multi-sensor setups with monocular geometry from a foundation model, and contributing a new annotated dataset and a real-world application. Strengths include the real-world fine-tuning data, cross-domain evaluation, an ablation over optimization steps T, and area estimation on real flight images. The immediate scientific contribution, however, is weakened because the central claim about depth-normal synergy is asserted rather than demonstrated; the comparison is confounded by pretraining and by an undescribed semi-supervised stage. The paper is therefore publishable only after major experimental revision.

major comments (5)
  1. [III-B-2, Eq. (2)] The central claim of depth-normal synergy is not tested. The SLZ flow consumes D_{t+1}, N_{t+1}, and SLZ_t only as concatenated inputs; no variant removes or corrupts these channels, no variant replaces them with RGB features, and no variant uses only the segmentation branch without geometric inputs. Without such an ablation, the cross-domain improvement in Table I could be attributed to the additional SLZ flow parameters or to the frozen encoder rather than to geometry. Please add at least one controlled condition (e.g., zeroed or random D/N inputs, or a variant consuming RGB features from the same encoder) and report the in-domain and cross-domain metrics for each.
  2. [Table I and Section IV-C] The comparison is confounded by pretraining scale: VisLanding starts from Metric3D V2 with an 800k-step pretraining and a DINOv2 ViT backbone, while SegFormer is initialized from ImageNet-1k and SafeUAV from HOV synthetic data. The cross-domain advantage may therefore reflect the stronger pretrained features, not the depth-normal mechanism. A baseline using the same Metric3D V2 backbone with an SLZ head but without depth/normal inputs, or equalizing pretraining across methods, is necessary to support the claim.
  3. [Section IV-B] The semi-supervised training stage (EMA teacher from 4k steps) is mentioned but never described in the Method, and no ablation measures its effect. This stage could substantially affect both in-domain and cross-domain results; as written, the contributions of fine-tuning, semi-supervised learning, and the SLZ branch are entangled. Please provide details of the teacher/student architecture and loss, and an ablation without this stage.
  4. [Section IV-C and IV-E] The in-domain performance is much lower than the baselines (76.61 mIoU vs. 93.72 for SegFormer in Table I), and the paper explains the cross-domain reversal only by assertion. Since the central claim is about generalization, please provide per-class results, error maps, or a direct analysis showing that the geometric inputs reduce domain-specific overfitting, rather than simply pointing to the overall cross-domain numbers.
  5. [Section IV-A-2] The cross-domain test labels are constructed by the authors from Semantic Drone categories, following [22] plus morphological dilation. All cross-domain numbers depend on this proxy, but no validation or sensitivity analysis is provided. Please report results under alternative labeling rules (e.g., without dilation, or with a small manually annotated subset) to show that the ranking in Table I is not an artifact of the labeling definition.
minor comments (6)
  1. [Section III-A-1] The binary map convention (0=safe, 1=unsafe) is the reverse of the common convention; please state it consistently with the class weights (safe=2) and the qualitative figures.
  2. [Section IV-B] 'Small version of the Metric3D V2 pretrained model' is not specified; report the exact model name, parameter count, and pretraining dataset.
  3. [Figure 4 and Section IV-D] The caption says 'both measured areas are approximately 34 m²', while the boxes show values from 17.59 to 36.34 m²; clarify which values are ground truth, which are predictions, and the temporal ordering.
  4. [Equation (9)] The area formula should state the assumptions (unit normal, locally planar surface) and define n_z(u,v) explicitly, including the sign of the normal.
  5. [Abstract and Section IV-C] The paper uses 'zero-shot' for the cross-domain SLZ evaluation, but the model is fine-tuned on WildUAV; please reserve the term for Metric3D V2's depth/normal predictions or redefine it explicitly.
  6. [Table I] Only a single run is reported; standard deviations or a note on the number of runs would help interpret the small differences between T=2, T=3, and T=4.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SLZ branch is trained with supervised ground truth and the area formula is a geometric projection, not a fitted quantity.

full rationale

VisLanding's central derivation is not circular. The SLZ branch is trained with a weighted cross-entropy loss (Eq. 7) against manually annotated WildUAV masks, and evaluated on held-out WildUAV images and on Semantic Drone images whose labels are constructed from semantic categories using the criteria of Loera-Ponce et al. This is a genuine transfer test rather than a restatement of training targets. The depth and normal channels in Eq. (2) are concatenated inputs, not quantities fitted to the SLZ labels; no equation redefines the SLZ output as a function of the evaluation labels. The landing-zone area estimate (Eq. 9) is a closed-form projection of predicted depth and normal pixels, with no parameters fit to the reported ~34 m^2 test areas. The only load-bearing external dependency is Metric3D V2's pretrained depth/normal backbone, which is cited as prior work by different authors and is a standard external pretrained model; even if its zero-shot generalization is not re-verified here, that is a verification gap, not an input-output equivalence. The absence of an RGB-only ablation weakens the mechanistic claim that depth-normal synergy causes the cross-domain gain, and the comparison is partly confounded by different pretraining, but these are correctness concerns, not circular reductions. Section IV-E acknowledges further limitations (dataset scale and real-time throughput), which again concern scope and deployment, not the derivation's internal consistency.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The ledger is dominated by hyperparameters of the pretrained model and the loss function, plus domain assumptions about safety definition and annotation quality. No new entities are invented. The largest epistemic load is carried by the assumption that Metric3D V2's geometry features transfer to the cross-domain test set, since the SLZ branch cannot be distinguished from a simple RGB-based head without an ablation.

free parameters (6)
  • lambda_1 (virtual normal loss weight) = 0.2
    Hand-chosen balancing coefficient in Eq. 3; not justified by sensitivity analysis.
  • lambda_2 (sequential depth loss weight) = 0.5
    Hand-chosen; affects the depth fine-tuning and thus the SLZ features.
  • lambda_3 (depth-normal consistency weight) = 0.01
    Hand-chosen; small value indicates marginal direct contribution.
  • gamma (temporal decay) = 0.9
    Decay factor for multi-stage losses in Eq. 3 and Eq. 7.
  • T (optimization steps) = 4
    Number of DPT iterations; Table I shows performance varies with T, so it is a tunable parameter.
  • class weight for safe class = 2
    Class imbalance weighting in L_safe; chosen without reported sensitivity.
assumptions (5)
  • domain assumption Safe landing zones are adequately captured by a binary mask based on flatness, absence of obstacles, and sufficient area.
    Adopted from the literature [2]; simplifies the task to binary segmentation and may mislabel regions that are flat but have hidden hazards.
  • domain assumption Metric3D V2 provides accurate zero-shot metric depth and surface normals in the UAV perspective and across domains.
    The SLZ branch relies on these geometric features; Section III-B-2 feeds D_{t+1} and N_{t+1} into the SLZ flow.
  • ad hoc to paper The Semantic Drone annotation rules from [22] are a valid proxy for landing safety.
    The authors map semantic classes to safety and apply morphological dilation (Section IV-A-2), which is a constructed criterion, not real safety labels.
  • domain assumption Camera intrinsics are known for test images.
    Required for canonical space transform and area back-projection in Eq. 8-9; not validated on Semantic Drone.
  • domain assumption Manual annotations of 1329 WildUAV images are consistent and correct.
    Annotated by authors with ISAT; no inter-annotator agreement measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VisLanding: Monocular 3D Perception for UAV Safe Landing via Depth-Normal Synergy." pith.science (2026). https://pith.science/paper/CIP7IKJU

@misc{pith2026250614525,
  author       = {Pith},
  title        = {Pith review of: VisLanding: Monocular 3D Perception for UAV Safe Landing via Depth-Normal Synergy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CIP7IKJU}},
  note         = {Machine review of arXiv:2506.14525}
}
read the original abstract

This paper presents VisLanding, a monocular 3D perception-based framework for safe UAV (Unmanned Aerial Vehicle) landing. Addressing the core challenge of autonomous UAV landing in complex and unknown environments, this study innovatively leverages the depth-normal synergy prediction capabilities of the Metric3D V2 model to construct an end-to-end safe landing zones (SLZ) estimation framework. By introducing a safe zone segmentation branch, we transform the landing zone estimation task into a binary semantic segmentation problem. The model is fine-tuned and annotated using the WildUAV dataset from a UAV perspective, while a cross-domain evaluation dataset is constructed to validate the model's robustness. Experimental results demonstrate that VisLanding significantly enhances the accuracy of safe zone identification through a depth-normal joint optimization mechanism, while retaining the zero-shot generalization advantages of Metric3D V2. The proposed method exhibits superior generalization and robustness in cross-domain testing compared to other approaches. Furthermore, it enables the estimation of landing zone area by integrating predicted depth and normal information, providing critical decision-making support for practical applications.

Figures

Figures reproduced from arXiv: 2506.14525 by the authors.

Figure 1
Figure 1. Top: Comparative performance of our method versus baseline approaches on both In-domain and Cross-domain test benchmarks for the SLZ estimation task. The proposed method demonstrates enhanced robustness and superior generalization capabilities across all key metrics. Bottom: Zero-shot testing results on real-world flight data. The green mask indicates pixels predicted as safe landing zones, blue bounding boxes highl… view at source ↗
Figure 2
Figure 2. Dual-flow optimization framework. (Left) Main pipeline: Input images are transformed into canonical space via camera parameters K, processed by DINOv2 [25] ViT [26] for multi-scale feature extraction (F1/4, F1/7, F1/14), then refined through T iterations in the Dense Prediction Transformer (DPT [27]). (Right) Dual optimization flows: The depth-normal flow (green) employs three ConvGRU blocks with projection heads fo… view at source ↗
Figure 3
Figure 3. Sample Examples from WildUAV and Semantic Drone Datasets. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Experimental results of safe landing zone area estimation in real-flight images show improved accuracy when closer to targets, with both measured [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative prediction results of our method compared with other approaches on two test benchmarks. The proposed method demonstrates significant [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Human-Inspired Neuro-Symbolic World Modeling and Logic Reasoning for Interpretable Safe UAV Landing Site Assessment

    cs.RO 2025-10 conditional novelty 6.0 of 10

    NeuroSymLand fuses a lightweight neural segmenter with hand-refined logic rules to rank safe UAV landing sites, beating four lightweight baselines in AirSim and on edge hardware while emitting proof traces.

Reference graph

Works this paper leans on

27 extracted references · 26 canonical work pages · cited by 1 Pith paper

  1. [22]

    Risk Assessment for Autonomous Landing in Urban Environments using Semantic Segmentation

    Loera-Ponce J A, Mercado-Ravell D A, Becerra-Dur ´an I, et al. Risk Assessment for Autonomous Landing in Urban Environments using Semantic Segmentation[J]. arXiv e-prints, 2024: arXiv: 2410.12988

  2. [1]

    Target localization for autonomous landing site detection: A review and preliminary result with static image photogrammetry[J]

    Subramanian J A, Asirvadam V S, Zulkifli S A B M, et al. Target localization for autonomous landing site detection: A review and preliminary result with static image photogrammetry[J]. Drones, 2023, 7(8): 509

  3. [2]

    Vision-based autonomous landing for the uav: A review[J]

    Xin L, Tang Z, Gai W, et al. Vision-based autonomous landing for the uav: A review[J]. Aerospace, 2022, 9(11): 634

  4. [3]

    Towards vision- based safe landing for an autonomous helicopter[J]

    Garcia-Pardo P J, Sukhatme G S, Montgomery J F. Towards vision- based safe landing for an autonomous helicopter[J]. Robotics and Autonomous Systems, 2002, 38(1): 19-29

  5. [4]

    Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation[J]

    Hu M, Yin W, Zhang C, et al. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

  6. [5]

    Wilduav: Monocular uav dataset for depth estimation tasks[C]//2021 IEEE 17th International Confer- ence on Intelligent Computer Communication and Processing (ICCP)

    Florea H, Miclea V C, Nedevschi S. Wilduav: Monocular uav dataset for depth estimation tasks[C]//2021 IEEE 17th International Confer- ence on Intelligent Computer Communication and Processing (ICCP). IEEE, 2021: 291-298

  7. [6]

    Zoedepth: Zero-shot transfer by com- bining relative and metric depth[J]

    Bhat S F, Birkl R, Wofk D, et al. Zoedepth: Zero-shot transfer by com- bining relative and metric depth[J]. arXiv preprint arXiv:2302.12288, 2023

  8. [7]

    Drone Dataset[DB/OL]

    ICG, Graz University of Technology. Drone Dataset[DB/OL]. http://dronedataset.icg.tugraz.at (2025-02-27)

Show all 27 references
  1. [8]

    UniDepth: Universal monocular metric depth estimation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Piccinelli L, Yang Y H, Sakaridis C, et al. UniDepth: Universal monocular metric depth estimation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 10106-10116

  2. [9]

    Towards zero-shot scale- aware monocular depth estimation[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision

    Guizilini V , Vasiljevic I, Chen D, et al. Towards zero-shot scale- aware monocular depth estimation[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023: 9233-9243

  3. [10]

    Depth anything v2[J]

    Yang L, Kang B, Huang Z, et al. Depth anything v2[J]. Advances in Neural Information Processing Systems, 2025, 37: 21875-21911

  4. [11]

    Repurposing diffusion-based image generators for monocular depth estimation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Ke B, Obukhov A, Huang S, et al. Repurposing diffusion-based image generators for monocular depth estimation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 9492-9502

  5. [12]

    Diffusiondepth: Diffusion denoising approach for monocular depth estimation[C]//European Conference on Com- puter Vision

    Duan Y , Guo X, Zhu Z. Diffusiondepth: Diffusion denoising approach for monocular depth estimation[C]//European Conference on Com- puter Vision. Cham: Springer Nature Switzerland, 2024: 432-449

  6. [13]

    Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

    Li Y , Zhang X, Chen D. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 1091-1100

  7. [14]

    Geometric and physical con- straints for drone-based head plane crowd density estimation[C]//2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Liu W, Lis K, Salzmann M, et al. Geometric and physical con- straints for drone-based head plane crowd density estimation[C]//2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2019: 244-249

  8. [15]

    Human crowd detection for drone flight safety using convolutional neural networks[C]//2017 25th European Signal Processing Conference (EUSIPCO)

    Tzelepi M, Tefas A. Human crowd detection for drone flight safety using convolutional neural networks[C]//2017 25th European Signal Processing Conference (EUSIPCO). IEEE, 2017: 743-747

  9. [16]

    Graph embedded convolutional neural networks in human crowd detection for drone flight safety[J]

    Tzelepi M, Tefas A. Graph embedded convolutional neural networks in human crowd detection for drone flight safety[J]. IEEE Transactions on Emerging Topics in Computational Intelligence, 2019, 5(2): 191- 204

  10. [17]

    SafeUA V: Learning to es- timate depth and safe landing areas for UA Vs from synthetic data[C]//Proceedings of the European Conference on Computer Vision (ECCV) Workshops

    Marcu A, Costea D, Licaret V , et al. SafeUA V: Learning to es- timate depth and safe landing areas for UA Vs from synthetic data[C]//Proceedings of the European Conference on Computer Vision (ECCV) Workshops. 2018: 0-0

  11. [18]

    Robust autonomous landing of UA Vs in non-cooperative environments based on comprehensive terrain un- derstanding[J]

    Chen L, Xiao Y , Yuan X, et al. Robust autonomous landing of UA Vs in non-cooperative environments based on comprehensive terrain un- derstanding[J]. Science China Information Sciences, 2022, 65(11): 212202

  12. [19]

    Safe landing zones detection for uavs using deep regression[C]//2022 19th conference on robots and vision (CRV)

    Abdollahzadeh S, Proulx P L, Allili M S, et al. Safe landing zones detection for uavs using deep regression[C]//2022 19th conference on robots and vision (CRV). IEEE, 2022: 213-218

  13. [20]

    YOLO-based terrain classification for uav safe landing zone detection[C]//2023 IEEE Region 10 Symposium (TENSYMP)

    Serrano K K D, Bandala A A. YOLO-based terrain classification for uav safe landing zone detection[C]//2023 IEEE Region 10 Symposium (TENSYMP). IEEE, 2023: 1-5

  14. [21]

    You only look once: Unified, real-time object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition

    Redmon J, Divvala S, Girshick R, et al. You only look once: Unified, real-time object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 779-788

  15. [23]

    SegFormer: Simple and efficient design for semantic segmentation with transformers[J]

    Xie E, Wang W, Yu Z, et al. SegFormer: Simple and efficient design for semantic segmentation with transformers[J]. Advances in neural information processing systems, 2021, 34: 12077-12090

  16. [24]

    (2024).ISAT with Segment Anything: An Interactive Semi-Automatic Annotation Tool

    Ji, S., & Zhang, H. (2024).ISAT with Segment Anything: An Interactive Semi-Automatic Annotation Tool. Retrieved from https://github.com/yatengLG/ISAT_with_segment_ anything. Updated on 2025-02-07, version 1.29

  17. [25]

    DINOv2: Learning Robust Visual Features without Supervision[J]

    Oquab M, Darcet T, Moutakanni T, et al. DINOv2: Learning Robust Visual Features without Supervision[J]. Transactions on Machine Learning Research Journal, 2024: 1-31

  18. [26]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale[C]//International Conference on Learning Representations

    Dosovitskiy A, Beyer L, Kolesnikov A, et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale[C]//International Conference on Learning Representations. 2020

  19. [27]

    Vision transformers for dense prediction[C]//Proceedings of the IEEE/CVF international conference on computer vision

    Ranftl R, Bochkovskiy A, Koltun V . Vision transformers for dense prediction[C]//Proceedings of the IEEE/CVF international conference on computer vision. 2021: 12179-12188

Pith tools

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