Computer Vision and Pattern Recognition · arXiv:2611.16984 v1 · submitted 12 Nov 2026 · 18 references
ElasticDino: Dense Semantic Features Using Elastic Deformations
The paper says
“Vision transformers (ViTs) serve as the backbone for countless vision tasks, including dense prediction tasks, such as segmentation and depth estimation, even though the quadratic complexity of self-attention severely limits feature resolution and induces spatial artifacts.”
more of the abstract
In this work, we introduce ElasticDino, a lightweight adapter which produces high-resolution feature maps, adhering to the fine details of the input image. As a key insight, we observe the consistency of artifacts across channels, and simultaneously resolve them to match the input image structure. Using a slim architecture based on spatial deformations, ElasticDino produces pixel-aligned, high-resolution feature maps efficiently. Leveraging deformations allows strictly preserving sharp semantics, and thereby avoiding the smoothness biases inherent to alternative approaches. Due to the absence of ground truth, we employ pixel-level semantic segmentation and depth estimation with linear heads to predict a set of downstream dense tasks, guiding the deformation process without leaking task-specific semantics into the model. We demonstrate superior accuracy over current methods on a range of popular dense task benchmarks, with better computational performance.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up. ElasticDino is a clean, well-executed paper that introduces a deformation-based upsampling method for ViT features, with strong empirical results and a clear architectural insight. This line is wrong
4 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 18 references say.
The paper says, at Section 3.2, Eq. (1)
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
The deformation field is defined as F = R(p_f(f) ⊕ p_I(I)) and applied coordinate-wise as f̂_{u,v} = f[F_{u,v}]. The paper states that this 'sampling from the original features' prevents hallucination and preserves semantics. However, the deformation is applied to features at the current resolution, which after the first upsampling step are bilinearly interpolated, not the original low-resolution features.
more of objection 1
The claim that the model 'diminishes hallucination hazards' by sampling from the original feature map is therefore only strictly true at the first deformation step; at later steps the features have already been interpolated. The authors should clarify what 'original features' means in this context and discuss whether the interpolation step re-introduces any of the smoothness biases the method aims to avoid.
The paper says, at Table 3, Cityscapes row
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
ElasticDino underperforms on Cityscapes (mIoU 61.33 vs. UPLiFT 65.02, Acc 93.44 vs. 94.34). The authors attribute this to small objects and provide a size-based analysis in Supplementary Figure 3. While this explanation is plausible, the gap is large (≈3.7 mIoU) and Cityscapes is a standard benchmark.
more of objection 2
The paper should report whether the same size-based pattern holds for the other methods (e.g., does UPLiFT also degrade on small objects, or does it handle them better?). Without this comparison, the claim that the limitation is inherent to the method rather than a weakness of the specific architecture remains partially speculative.
The paper says, at Section 3.3, multi-scale loss
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The multi-scale loss L_ms sums losses over all stages i, all heads j, and all resolutions r_k. This means gradients flow into every intermediate feature map. The paper states this 'greatly stabilizes and speeds up the learning' but does not ablate this design choice.
more of objection 3
Since the deformation blocks are already lightweight, it is unclear whether the multi-scale loss is necessary or whether a loss only at the final resolution would suffice. An ablation comparing single-scale vs. multi-scale loss would strengthen the paper.
The paper says, at Section 4.2, Table 5
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
The ablation study evaluates depth estimation on a held-out test set from Hypersim. The 'w/o DEFormations' variant shows a catastrophic drop (δ1 from 58.35% to 26.80%). This is a strong result, but the paper does not specify how the 'w/o DEFormations' variant is constructed.
more of objection 4
Is it simply the same architecture with the deformation blocks removed (i.e., only bilinear upsampling and convolutions)? If so, the comparison conflates the removal of deformations with a change in the number of parameters and the receptive field. The authors should clarify the exact architecture of this ablation and whether the number of parameters was matched.
6 minor objections
The paper says, at Abstract
Pith says: minor objection 1 of 6
The abstract states 'the quadratic complexity of self-attention severely limits feature resolution and induces spatial artifacts.' The quadratic complexity is a computational constraint, not a direct cause of artifacts; the artifacts stem from patchification and positional encoding. Consider rephrasing for precision.
The paper says, at Section 3.1
Pith says: minor objection 2 of 6
The cross-channel consistency analysis uses Spearman correlation of gradient magnitudes. The reported values (0.29 vs. 0.11 for DINOv2) are modest in absolute terms. The paper should clarify whether these differences are statistically significant (e.g., with confidence intervals or p-values).
show 4 more
The paper says, at Table 1
Pith says: minor objection 3 of 6
The parameter count for ElasticDino is listed as 27,095K (27M), which is substantially larger than competitors (e.g., JAFAR 628K). The paper notes that the model is faster and more memory-efficient despite this, but the large parameter count should be discussed in context: are the extra parameters in the task heads (which are discarded after training) or in the upsampler itself? If the latter, the claim of 'lightweight' should be qualified.
The paper says, at Supplementary Table 3
Pith says: minor objection 4 of 6
The no-depth and no-seg variants show strong performance on held-out tasks, supporting the claim of no task leakage. However, the no-depth variant achieves δ1=60.5% vs. the full model's 64.14%, and the no-seg variant achieves 64.8%. The no-seg variant actually outperforms the full model on depth RMSE (0.56 vs. 0.57). This non-monotonic behavior is interesting and deserves a brief comment.
The paper says, at Section 5.1
Pith says: minor objection 5 of 6
The limitation about small holes is stated but not quantified. The paper should provide a quantitative measure (e.g., performance vs. object size on a standard benchmark) to complement the qualitative inset figure.
The paper says, at References
Pith says: minor objection 6 of 6
Reference [6] (Chollet, Xception) is cited for depthwise separable convolutions, which is appropriate. However, the paper uses 'depthwise-separable' and 'depth-wise separable' interchangeably; please standardize.
What would settle it
Verify that the deformation fields are indeed smooth and invertible (or at least not collapsing) by computing the Jacobian determinant of the predicted flow field on a held-out set of 100 images; if more than 5% of pixels have a non-positive determinant, the deformation may introduce folding artifacts that could affect feature semantics.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 13 Nov 2026
- Reader holds up
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee minor revision: 4 major objections, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The paper introduces ElasticDino, a lightweight upsampling method for ViT features that uses elastic deformations rather than attention or bilateral filtering. The core insight—that ViT artifacts are cross-channel consistent and can be corrected with a single deformation field—is well supported by a simple correlation analysis (Section 3.1). The architecture is clean: alternating bilinear upsampling and learned deformations, guided by the input image, with no feature-value editing. Training uses 1x1 heads on downstream tasks (segmentation, depth, etc.) to provide supervision without leaking task-specific semantics into the features. The held-out task ablations (Table 5, Supplementary Table 3) back this claim convincingly.
The empirical work is extensive and honestly reported. The method outperforms prior upsampling approaches (JAFAR, AnyUp, LoftUp, UPLiFT) on most benchmarks, with lower compute and memory (Table 1). The authors also acknowledge the main limitation—poor performance on small objects and holes—and provide a size-based analysis (Supplementary Figure 3) that explains the Cityscapes results. The ablation study is thorough and includes a leave-one-task-out design.
Soft spots: (1) The training relies on pseudo-labels from Mask2Former and DepthAnything. The authors argue this is acceptable because the deformation architecture prevents task leakage, and the held-out experiments support this. I find the argument plausible but not airtight—the pseudo-labels could still bias the deformation field in subtle ways. (2) The method is evaluated only on DINOv2, DINOv3, and CLIP backbones. It would be useful to see results on other ViT families (e.g., MAE, DeiT) to assess generality. (3) The paper claims "superior accuracy" on all benchmarks, but Cityscapes is a clear exception (the method is second-best there). The authors address this honestly, but the claim in the abstract is slightly overstated.
Overall, this is a solid, well-written paper with a genuine architectural contribution. The deformation-based approach is novel and the empirical support is strong. I would bring this to a reading group and would cite it. It deserves a serious peer review.
Recommendation: Accept with minor revisions. The authors should add a discussion of pseudo-label bias and ideally test on one additional backbone family.
The strongest objection
The paper's central claim is that ElasticDino produces high-resolution, artifact-corrected feature maps that outperform existing upsampling methods on dense prediction tasks. The weakest assumption identified by the reader—that training with 1x1 heads on pseudo-labels might leak task-specific semantics—is explicitly addressed by the authors: they argue that the deformation-based architecture prevents feature-value editing, and they provide ablation studies (Table 5, Supplementary Table 3) showing that held-out tasks (depth, segmentation) remain strong when those tasks are omitted from training. The paper also acknowledges its main limitation (small objects/holes) and provides a size-based analysis (Supplementary Figure 3). The empirical results are extensive, covering multiple backbones (DINOv2, DINOv3, CLIP), multiple datasets, and multiple tasks, with consistent improvements. The computational efficiency claims are supported by Table 1. I find no unresolved load-bearing concern that would undermine the central claim.
The reader's weakest_assumption focuses on the risk of task-specific semantics leaking into the upsampled features due to training with pseudo-labels. The paper directly addresses this concern: (1) the use of 1x1 convolutional heads prevents spatial mixing, (2) the deformation-based architecture samples from original features rather than editing them, and (3) ablation studies (Table 5, Supplementary Table 3) show that omitting depth or segmentation tasks during training does not degrade performance on those held-out tasks. The reader's concern is valid in principle but is mitigated by the paper's design and evidence. My partial agreement reflects that while the concern is reasonable, the paper provides sufficient counter-evidence. The only potential concern I could identify is whether the deformation fields might introduce folding artifacts (non-invertible mappings), but this is a technical detail that does not threaten the central claim. The paper's own limitation about small objects is acknowledged and analyzed. Therefore, I recommend no change to the ACCEPT verdict.
The referee's report
If the results hold, ElasticDino offers a practical and principled solution to a known limitation of ViT-based dense prediction: the low resolution and artifacts of patch-level features. The deformation-based approach is conceptually clean—it avoids feature-value editing and the associated risk of task leakage—and the empirical gains are consistent across multiple backbones, tasks, and datasets. The computational efficiency (Table 1) is a genuine advantage over attention-based competitors. The paper also provides a useful analysis of cross-channel artifact consistency (Section 3.1) that justifies the single-deformation-field design. The main limitation (small objects/holes) is acknowledged and analyzed. Overall, the work represents a solid contribution to the feature-upsampling literature.
Pith's reading of the paper (a machine's summary, not the authors')
Vision transformers like DINO produce rich semantic features but at low resolution with artifacts due to patchification and positional encodings. ElasticDino is a lightweight adapter that upscales these features to high resolution while correcting artifacts. The key insight is that artifacts are consistent across feature channels, so a single deformation field can correct all channels simultaneously. The method alternates between bilinear upsampling and applying learned deformation fields that warp the features to match the structure of the input image. Because there is no ground truth for upscaled features, ElasticDino is trained using a set of simple 1x1 convolutional heads that predict various dense maps (RGB, depth, segmentation, normals, intrinsic decomposition) from the upscaled features. The 1x1 receptive field forces the heads to use only the semantics at each pixel, preventing the upscaler from encoding task-specific information. Experiments show that ElasticDino outperforms previous upsampling methods (FeatUp, JAFAR, AnyUp, LoftUp, UPLiFT) on semantic segmentation, open-vocabulary segmentation, and depth estimation benchmarks, while being faster and using less memory.
Core claim. ElasticDino produces high-resolution feature maps, adhering to the fine details of the input image, and demonstrates superior accuracy over current methods on a range of popular dense task benchmarks, with better computational performance (Section 1, Abstract).
The premise it rests on. The assumption that training with 1x1 convolutional heads on downstream tasks using pseudo-labels (from Mask2Former and DepthAnything) provides sufficient guidance to learn a general upsampler without leaking task-specific semantics. This is stated in Section 3.2: 'we employ pixel-level linear heads to predict a set of downstream dense tasks, guiding the deformation process without leaking task specific semantics into the model.' The reliance on pseudo-labels means the upsampler is indirectly trained on potentially noisy targets.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Vision transformers (ViTs) serve as the backbone for countless vision tasks, including dense prediction tasks, such as segmentation and depth estimation, even though the quadratic complexity of self-attention severely limits feature resolution and induces spatial artifacts. In this work, we introduce ElasticDino, a lightweight adapter which produces high-resolution feature maps, adhering to the fine details of the input image. As a key insight, we observe the consistency of artifacts across channels, and simultaneously resolve them to match the input image structure. Using a slim architecture based on spatial deformations, ElasticDino produces pixel-aligned, high-resolution feature maps efficiently. Leveraging deformations allows strictly preserving sharp semantics, and thereby avoiding the smoothness biases inherent to alternative approaches. Due to the absence of ground truth, we employ pixel-level semantic segmentation and depth estimation with linear heads to predict a set of downstream dense tasks, guiding the deformation process without leaking task-specific semantics into the model. We demonstrate superior accuracy over current methods on a range of popular dense task benchmarks, with better computational performance.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-13
The authors' main claim, as the classifier picked it out 0.99
“We demonstrate superior accuracy over current methods on a range of popular dense task benchmarks, with better computational performance.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.44 · human participants 0.02 · controlled comparison 0.56 · randomized assignment 0.01 · states quantitative result 0.05 · compares with prior results 0.93 · mentions code or data availability 0.02
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Vision transformers (ViTs) serve as the backbone for countless vision tasks, including dense prediction tasks, such as segmentation and depth estimation, even though the quadratic complexity of self-attention severely limits feature resolution and induces spatial artifacts. | background 0.99 | 0.07 | 0.03 | asserted 1.00 |
| In this work, we introduce ElasticDino, a lightweight adapter which produces high-resolution feature maps, adhering to the fine details of the input image. | contribution 0.99 | 0.89 | 0.15 | asserted 1.00 |
| As a key insight, we observe the consistency of artifacts across channels, and simultaneously resolve them to match the input image structure. | method 0.45 | 0.87 | 0.10 | asserted 1.00 |
| Using a slim architecture based on spatial deformations, ElasticDino produces pixel-aligned, high-resolution feature maps efficiently. | method 0.85 | 0.89 | 0.17 | asserted 1.00 |
| Leveraging deformations allows strictly preserving sharp semantics, and thereby avoiding the smoothness biases inherent to alternative approaches. | interpretation 0.56 | 0.58 | 0.13 | asserted 0.99 |
| Due to the absence of ground truth, we employ pixel-level semantic segmentation and depth estimation with linear heads to predict a set of downstream dense tasks, guiding the deformation process without leaking task-specific semantics into the model. | method 1.00 | 0.93 | 0.05 | asserted 0.98 |
| We demonstrate superior accuracy over current methods on a range of popular dense task benchmarks, with better computational performance. | result 1.00 | 0.96 | 0.74 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-13: 1 request, 5,602 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (13)
Figures are the authors', reproduced from the paper.
References (18)
- [41] Simon Baker, Daniel Scharstein, James P Lewis, Stefan Roth, Michael J Black, and Richard Szeliski. A database and evaluation methodology for optical flow. International journal of computer vision, 92(1):1–31, 2011 on Pith
- [42] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexan- der Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. arXiv, 2021 on Pith
- [43] Franc ¸ois Chollet. Xception: Deep learning with depthwise separable convolutions, 2017 on Pith
- [44] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding, 2016 on Pith
- [45] Paul Couairon, Loick Chambon, Louis Serrano, Jean- Emmanuel Haugeard, Matthieu Cord, and Nicolas Thome. Jafar: Jack up any feature at any resolution, 2025 on Pith
- [46] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009 on Pith
- [47] Haiwen Huang, Anpei Chen, V olodymyr Havrylov, Andreas Geiger, and Dan Zhang. Loftup: Learning a coordinate- based feature upsampler for vision foundation models.arXiv preprint arXiv:2504.14032, 2025 on PithPith's reviewarXiv
- [48] Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation, 2022 on Pith
- [49] Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700, 2019 on PithPith's reviewarXiv
- [50] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Doll ´ar. Microsoft coco: Common objects in context, 2015 on Pith
- [51] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019 on Pith
- [52] 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 on PithPith's reviewarXiv
- [53] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic syn- thetic dataset for holistic indoor scene understanding, 2021 on Pith
- [54] Oriane Sim ´eoni, Huy V . V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timoth´ee Darcet, Th´eo Moutakanni, Leonel Sen... on Pith
- [55] Matthew Walmer, Saksham Suri, Anirud Aggarwal, and Abhinav Shrivastava. Uplift: Efficient pixel-dense fea- ture upsampling with local attenders. arXiv preprint arXiv:2601.17950, 2026 on PithPith's reviewarXiv
- [56] Thomas Wimmer, Prune Truong, Marie-Julie Rakotosaona, Michael Oechsle, Federico Tombari, Bernt Schiele, and Jan Eric Lenssen. Anyup: Universal feature upsampling. arXiv preprint arXiv:2510.12764, 2025 on PitharXiv
- [57] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data, 2024 on Pith
- [58] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision (ECCV), 2022. 25 on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment