pith. sign in

arxiv: 2606.22749 · v1 · pith:LWEM5HS2new · submitted 2026-06-22 · 💻 cs.CV

RaysUp: Ultra-light Universal Feature Upsampling via Geometry-Aware Ray Representation

Yuchuan Ding , Linfei Li , Lin Zhang , Ying Shen This is my paper

Pith reviewed 2026-06-26 09:10 UTC · model grok-4.3

classification 💻 cs.CV
keywords feature upsamplingvision foundation modelsray representationPlucker coordinatesdense predictionlightweight modelcross-attentiongeometry-aware attention
0
0 comments X

The pith

Feature upsampling from vision foundation models can be made ultra-lightweight and resolution-flexible by operating in a geometry-aware ray domain.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that lifting low-resolution VFM features into a ray domain with explicit 3D geometric priors enables high-fidelity reconstruction at arbitrary scales without task-specific retraining. This would matter if true because existing methods either lose semantic detail through simple interpolation or demand heavy architectures and per-VFM tuning that limit scalability. RaysUp uses Plucker ray coordinates for positional encoding, decoupled guidance, flexible cross-attention, and geometry-aware neighborhood attention to achieve the reconstruction. Experiments across dense prediction tasks show the approach reaches state-of-the-art results with substantially lower parameter counts and faster inference than prior universal upsamplers.

Core claim

RaysUp is an ultra-lightweight, task-agnostic and VFM-agnostic framework that reconstructs high-resolution feature maps at arbitrary resolutions by lifting feature reconstruction into a geometry-aware ray domain. It introduces Ray Positional Encoding with 6D Plucker ray coordinates to inject implicit 3D geometric priors, a Spatially Decoupled Guidance Encoder for direction-aware guidance, Any-Resolution Cross-Attention for resolution-flexible reconstruction, and a Geometry-Aware Neighborhood Attention module for content-adaptive bilateral aggregation while preserving geometric consistency. Extensive experiments demonstrate that this yields state-of-the-art performance while using only 16% of

What carries the argument

Geometry-aware ray representation that encodes features via 6D Plucker ray coordinates to inject 3D geometric priors into the upsampling process.

If this is right

  • Arbitrary-resolution feature maps can be produced without retraining the underlying vision foundation model.
  • Semantic fidelity is maintained across multiple dense prediction tasks using a single set of weights.
  • Parameter count and inference time drop substantially relative to prior universal upsamplers while accuracy improves.
  • The same architecture works for any VFM output without architecture-specific adjustments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The explicit 3D priors in the ray domain could reduce the need for learned positional encodings in other 2D vision modules.
  • Lightweight ray-based reconstruction may enable on-device high-resolution inference for mobile dense prediction pipelines.
  • The decoupled guidance and neighborhood attention patterns could be adapted to video or multi-view feature fusion.

Load-bearing premise

The ray-domain components deliver content-adaptive reconstruction that preserves semantic fidelity without VFM-specific retraining or task-dependent tuning.

What would settle it

Applying RaysUp to a previously unseen VFM on a dense prediction task and measuring whether semantic fidelity drops below that of a retrained baseline in fine-grained regions.

Figures

Figures reproduced from arXiv: 2606.22749 by Linfei Li, Lin Zhang, Ying Shen, Yuchuan Ding.

Figure 1
Figure 1. Figure 1: We propose RaysUp, an (a) ultra-lightweight, (b) task-agnostic, and (c) VFM￾agnostic upsampling framework, capable of upsampling backbone features to arbitrary resolutions while preserving (d) high semantic fidelity and geometric consistency. to their large-scale pretraining on diverse datasets, which endows them with high￾level semantic representations and strong generalization capability. These models ar… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of RaysUp. Given an image I and a low-resolution VFM fea￾ture map F lr, RaysUp reconstructs a high-resolution feature map F hr at an arbi￾trary target resolution. A lightweight Spatially Decoupled Guidance Encoder first ex￾tracts direction-aware guidance features Fg, which are adaptively pooled to generate target-resolution queries Qg and VFM-resolution keys Kg. After RayPE, the resulting geometry… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative Results for Task-agnostic Performance. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of convolutional kernel weights. [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Attention bias map from the query pixel (×). B Implicit Geometric Injection in RayPE RoPE is defined on a planar isotropic 2D Euclidean grid, where the positional phase is linearly associated with pixel coordinates (i, j): θRoPE(i, j) = [iω, jω] T (ω denotes the frequency). In contrast, RayPE implicitly encodes 3D spatial positions using normalized camera rays. Assuming an identity camera extrinsic matrix,… view at source ↗
Figure 6
Figure 6. Figure 6: Additional pca visualizations on ImageNet. [PITH_FULL_IMAGE:figures/full_fig_p026_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional depth estimation visualizations on NYUv2. [PITH_FULL_IMAGE:figures/full_fig_p027_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional surface normal estimation visualizations on NYUv2. [PITH_FULL_IMAGE:figures/full_fig_p028_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional semantic segmentation visualizations on VOC. [PITH_FULL_IMAGE:figures/full_fig_p029_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional video semantic segmentation visualizations on the [PITH_FULL_IMAGE:figures/full_fig_p030_10.png] view at source ↗
read the original abstract

Pre-trained Vision Foundation Models (VFMs) have become central to modern computer vision due to their powerful semantic representations and strong generalization ability. However, their patchified or pooled outputs are inherently low-resolution, limiting their effectiveness in tasks requiring fine-grained, pixel-level reasoning. Existing feature upsampling approaches either degrade semantic fidelity or rely on VFM-specific retraining and heavy architectures, hindering efficiency and scalability. To address these challenges, we propose RaysUp, an ultra-lightweight, task-agnostic, and VFM-agnostic feature upsampling framework that reconstructs high-resolution feature maps at arbitrary resolutions. Unlike conventional 2D interpolation or attention-based schemes, RaysUp lifts feature reconstruction into a geometry-aware ray domain. Specifically, we introduce a Spatially Decoupled Guidance Encoder for direction-aware guidance encoding, an Any-Resolution Cross-Attention mechanism for resolution-flexible reconstruction, and a novel Ray Positional Encoding (RayPE) that injects implicit 3D geometric priors via 6D Plucker ray coordinates. Finally, a Geometry-Aware Neighborhood Attention module further ensures content-adaptive bilateral aggregation while preserving geometric consistency. Extensive experiments across diverse dense prediction tasks demonstrate that RaysUp achieves state-of-the-art performance while using only 16% of the parameters of AnyUp and delivering approximately 7x faster inference. These results highlight a substantially improved accuracy-efficiency trade-off and establish RaysUp as a practical and scalable solution for universal feature upsampling. Code is available at https://github.com/MAP-RaysUp/RaysUp.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces RaysUp, an ultra-lightweight feature upsampling framework for pre-trained Vision Foundation Models (VFMs). It lifts low-resolution VFM outputs into a geometry-aware ray domain using a novel Ray Positional Encoding (RayPE) based on 6D Plucker coordinates, a Spatially Decoupled Guidance Encoder, Any-Resolution Cross-Attention, and Geometry-Aware Neighborhood Attention. The central claim is that this yields state-of-the-art accuracy-efficiency trade-offs across dense prediction tasks while using only 16% of AnyUp's parameters and running approximately 7x faster, all in a task- and VFM-agnostic manner without retraining.

Significance. If the performance claims and universality hold, RaysUp would offer a practical, scalable solution for high-resolution feature reconstruction in dense prediction, improving the accuracy-efficiency frontier without VFM-specific tuning. The introduction of ray-domain geometric priors is a potentially novel direction, though its independence from fitted parameters is not yet demonstrated.

major comments (1)
  1. [Abstract / RayPE description] Abstract and RayPE section: The 6D Plucker coordinates (direction d and moment m = o × d) in RayPE presuppose known camera intrinsics K and poses to compute normalized directions and moments from pixel coordinates. No mechanism is described for obtaining K or poses from arbitrary VFM feature maps alone, which directly contradicts the VFM-agnostic and task-agnostic claims; if the method silently assumes calibration or falls back to 2D approximations, the geometry-aware prior is not actually realized.
minor comments (1)
  1. [Abstract] The abstract asserts SOTA performance, 16% parameter count, and 7x speedup, but the provided text contains no quantitative tables, ablation studies, error bars, or dataset details, preventing direct evaluation of the central claims.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for identifying this point on the geometric encoding. We respond to the major comment below.

read point-by-point responses

  1. Referee: [Abstract / RayPE description] Abstract and RayPE section: The 6D Plucker coordinates (direction d and moment m = o × d) in RayPE presuppose known camera intrinsics K and poses to compute normalized directions and moments from pixel coordinates. No mechanism is described for obtaining K or poses from arbitrary VFM feature maps alone, which directly contradicts the VFM-agnostic and task-agnostic claims; if the method silently assumes calibration or falls back to 2D approximations, the geometry-aware prior is not actually realized.

    Authors: We agree that the manuscript should more explicitly state how camera parameters are obtained. RaysUp is applied to the original input images that accompany every VFM feature map in the evaluated dense-prediction benchmarks; camera intrinsics K and poses are taken from the dataset metadata (or estimated via standard calibration when absent) to compute the 6D Plucker rays before feature upsampling. This preprocessing step is independent of the VFM weights, preserving the VFM-agnostic property. In the revised manuscript we will expand the RayPE section with the explicit computation pipeline, pseudocode, and a short discussion of the assumption, thereby removing any ambiguity while leaving the core claims unchanged. revision: yes

Circularity Check

0 steps flagged

No significant circularity; architectural modules are independent contributions

full rationale

The paper presents RaysUp as a new framework with novel components (RayPE via 6D Plucker coordinates, Spatially Decoupled Guidance Encoder, Any-Resolution Cross-Attention, Geometry-Aware Neighborhood Attention) introduced as independent design choices for lifting feature reconstruction into a ray domain. No equations, predictions, or central claims in the abstract or described architecture reduce by construction to fitted parameters, self-citations, or renamed inputs. The derivation chain for task-agnostic upsampling relies on these explicitly defined modules without self-definitional loops or load-bearing self-citations. The method is self-contained against external benchmarks as a proposed architecture rather than a derived result from prior fitted values.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The method rests on the domain assumption that 6D Plucker ray coordinates supply useful implicit 3D geometric priors for 2D feature reconstruction; no free parameters or invented entities beyond the named architectural modules are quantified in the abstract.

axioms (1)
  • domain assumption Plucker ray coordinates inject implicit 3D geometric priors suitable for 2D feature map reconstruction
    Invoked in the description of Ray Positional Encoding (RayPE)
invented entities (1)
  • Ray Positional Encoding (RayPE) no independent evidence
    purpose: Inject implicit 3D geometric priors via 6D Plucker coordinates
    New encoding module introduced to replace conventional positional encodings

pith-pipeline@v0.9.1-grok · 5815 in / 1224 out tokens · 24736 ms · 2026-06-26T09:10:02.063081+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

63 extracted references · 5 canonical work pages · 5 internal anchors

  1. [1]

    In: CVPR

    Asim, M., Wewer, C., Wimmer, T., Schiele, B., Lenssen, J.E.: MEt3R: Measuring Multi-View Consistency in Generated Images. In: CVPR. pp. 6034–6044 (2025)

    2025

  2. [2]

    In: ICCV (2021)

    Bae, G., Budvytis, I., Cipolla, R.: Estimating and Exploiting the Aleatoric Uncer- tainty in Surface Normal Estimation. In: ICCV (2021)

    2021

  3. [3]

    In: ICCV

    Barsellotti,L.,Bianchi,L.,Messina,N.,Carrara,F.,Cornia,M.,Baraldi,L.,Falchi, F., Cucchiara, R.: Talking to DINO: Bridging Self-Supervised Vision Backbones with Language for Open-Vocabulary Segmentation. In: ICCV. pp. 22025–22035 (2025)

    2025

  4. [4]

    In: CVPR (2020)

    Bhat, S.F., Alhashim, I., Wonka, P.: AdaBins: Depth Estimation Using Adaptive Bins. In: CVPR (2020)

    2020

  5. [5]

    Perception Encoder: The best visual embeddings are not at the output of the network

    Bolya, D., Huang, P.Y., Sun, P., Cho, J.H., Madotto, A., Wei, C., Ma, T., Zhi, J., Rajasegaran, J., Rasheed, H., et al.: Perception Encoder: The Best Visual Em- beddings Are Not at the Output of the Network. arXiv preprint arXiv:2504.13181 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    In: CVPR

    Caesar, H., Uijlings, J., Ferrari, V.: COCO-Stuff: Thing and Stuff Classes in Con- text. In: CVPR. pp. 1209–1218 (2018)

    2018

  7. [7]

    In: ICCV

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging Properties in Self-Supervised Vision Transformers. In: ICCV. pp. 9650–9660 (2021) 16 Ding et al

    2021

  8. [8]

    In: CVPR

    Chen, S., Guo, H., Zhu, S., Zhang, F., Huang, Z., Feng, J., Kang, B.: Video Depth Anything: Consistent Depth Estimation for Super-Long Videos. In: CVPR. pp. 22831–22840 (2025)

    2025

  9. [9]

    In: NeurIPS (2025)

    Chuang, Y.S., Li, Y., Wang, D., Yeh, C.F., Lyu, K., Raghavendra, R., Glass, J.R., HUANG, L., Weston, J.E., Zettlemoyer, L., Chen, X., Liu, Z., Xie, S., tau Yih, W., Li, S.W., Xu, H.: Meta CLIP 2: A Worldwide Scaling Recipe. In: NeurIPS (2025)

    2025

  10. [10]

    In: CVPR

    Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The Cityscapes Dataset for Semantic Urban Scene Understanding. In: CVPR. pp. 3213–3223 (2016)

    2016

  11. [11]

    In: NeurIPS (2025)

    Couairon, P., Chambon, L., Serrano, L., HAUGEARD, J.E., Cord, M., THOME, N.: JAFAR: Jack up Any Feature at Any Resolution. In: NeurIPS (2025)

    2025

  12. [12]

    In: CVPR

    Dai, Y., Lu, H., Shen, C.: Learning Affinity-Aware Upsampling for Deep Image Matting. In: CVPR. pp. 6841–6850 (2021)

    2021

  13. [13]

    In: CVPR

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A Large- Scale Hierarchical Image Database. In: CVPR. pp. 248–255 (2009)

    2009

  14. [14]

    In: ICCV

    Ding, X., Guo, Y., Ding, G., Han, J.: ACNet: Strengthening the Kernel Skeletons for Powerful CNN via Asymmetric Convolution Blocks. In: ICCV. pp. 1911–1920 (2019)

    1911

  15. [15]

    In: CVPR

    Ding, X., Zhang, X., Ma, N., Han, J., Ding, G., Sun, J.: RepVGG: Making VGG- style ConvNets Great Again. In: CVPR. pp. 13733–13742 (2021)

    2021

  16. [16]

    Journal of Applied Meteorology18(8), 1016–1022 (1979)

    Duchon, C.E.: Lanczos Filtering in One and Two Dimensions. Journal of Applied Meteorology18(8), 1016–1022 (1979)

    1979

  17. [17]

    In: CVPR

    El Banani, M., Raj, A., Maninis, K.K., Kar, A., Li, Y., Rubinstein, M., Sun, D., Guibas, L., Johnson, J., Jampani, V.: Probing the 3D Awareness of Visual Foundation Models. In: CVPR. pp. 21795–21806 (2024)

    2024

  18. [18]

    Neural Networks107, 3–11 (2018)

    Elfwing, S., Uchibe, E., Doya, K.: Sigmoid-Weighted Linear Units for Neural Net- work Function Approximation in Reinforcement Learning. Neural Networks107, 3–11 (2018)

    2018

  19. [19]

    IJCV111(1), 98–136 (2015)

    Everingham, M., Eslami, S.M., Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The Pascal Visual Object Classes Challenge: A Retrospective. IJCV111(1), 98–136 (2015)

    2015

  20. [20]

    In: ICLR (2024)

    Fu, S., Hamilton, M., Brandt, L.E., Feldmann, A., Zhang, Z., Freeman, W.T.: FeatUp: A Model-Agnostic Framework for Features at Any Resolution. In: ICLR (2024)

    2024

  21. [21]

    In: CVPR

    Hassani,A.,Walton,S.,Li,J.,Li,S.,Shi,H.:NeighborhoodAttentionTransformer. In: CVPR. pp. 6185–6194 (2023)

    2023

  22. [22]

    In: CVPR

    He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked Autoencoders Are Scalable Vision Learners. In: CVPR. pp. 16000–16009 (2022)

    2022

  23. [23]

    In: ICCV (2025)

    Huang, H., Chen, A., Havrylov, V., Geiger, A., Zhang, D.: LoftUp: Learning a Coordinate-Based Feature Upsampler for Vision Foundation Models. In: ICCV (2025)

    2025

  24. [24]

    In: CVPR

    Jose, C., Moutakanni, T., Kang, D., Baldassarre, F., Darcet, T., Xu, H., Li, D., Szafraniec, M., Ramamonjisoa, M., Oquab, M., et al.: DINOv2 Meets Text: A Unified Framework for Image- and Pixel-Level Vision-Language Alignment. In: CVPR. pp. 24905–24916 (2025)

    2025

  25. [25]

    Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Regis- tration

    Jun-Seong, K., Kim, G., Yu-Ji, K., Wang, Y.C.F., Choe, J., Oh, T.H.: Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language Embedding Regis- tration. In: CVPR. pp. 14137–14146 (2025)

    2025

  26. [26]

    In: ICCV

    Kerr, J., Kim, C.M., Goldberg, K., Kanazawa, A., Tancik, M.: LERF: Language Embedded Radiance Fields. In: ICCV. pp. 19729–19739 (2023) RaysUp: Ultra-light Universal Feature Upsampling 17

    2023

  27. [27]

    In: ICCV

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment Anything. In: ICCV. pp. 4015–4026 (2023)

    2023

  28. [28]

    ACM Trans

    Kopf, J., Cohen, M.F., Lischinski, D., Uyttendaele, M.: Joint Bilateral Upsampling. ACM Trans. Graph.26(3), 96 (2007)

    2007

  29. [29]

    In: ECCV

    Lan, M., Chen, C., Ke, Y., Wang, X., Feng, L., Zhang, W.: ProxyCLIP: Proxy Attention Improves CLIP for Open-Vocabulary Segmentation. In: ECCV. pp. 70– 88 (2024)

    2024

  30. [30]

    In: ACM MM

    Li, L., Zhang, L., Wang, Z., Shen, Y.: GS3LAM: Gaussian Semantic Splatting SLAM. In: ACM MM. p. 3019–3027 (2024)

    2024

  31. [31]

    AAAI39(23), 24458–24466 (2025)

    Li, L., Zhang, L., Wang, Z., Zhang, F., Li, Z., Shen, Y.: Representing sounds as neural amplitude fields: A benchmark of coordinate-mlps and a fourier kolmogorov- arnold framework. AAAI39(23), 24458–24466 (2025)

    2025

  32. [32]

    In: ICLR (2026)

    Lin, H., Chen, S., Liew, J.H., Chen, D.Y., Li, Z., Zhao, Y., Peng, S., Guo, H., Zhou, X., Shi, G., Feng, J., Kang, B.: Depth Anything 3: Recovering the Visual Space from Any Views. In: ICLR (2026)

    2026

  33. [33]

    In: ICLR (2019)

    Loshchilov, I., Hutter, F.: Decoupled Weight Decay Regularization. In: ICLR (2019)

    2019

  34. [34]

    IEEE TPAMI44(1), 242–255 (2022)

    Lu, H., Dai, Y., Shen, C., Xu, S.: Index Networks. IEEE TPAMI44(1), 242–255 (2022)

    2022

  35. [35]

    In: ECCV

    Lu, H., Liu, W., Fu, H., Cao, Z.: FADE: Fusing the Assets of Decoder and Encoder for Task-Agnostic Upsampling. In: ECCV. pp. 231–247 (2022)

    2022

  36. [36]

    In: NeurIPS (2022)

    Lu, H., Liu, W., Ye, Z., Fu, H., Liu, Y., Cao, Z.: SAPA: Similarity-Aware Point Affiliation for Feature Upsampling. In: NeurIPS (2022)

    2022

  37. [37]

    College of the Redwoods 45(1), 1049–1060 (1998)

    McKinley, S., Levine, M.: Cubic Spline Interpolation. College of the Redwoods 45(1), 1049–1060 (1998)

    1998

  38. [38]

    In: ECCV

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In: ECCV. pp. 405–421 (2020)

    2020

  39. [39]

    TMLR (2024)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P., Li, S., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jégou, H., Mairal, J., Labatut, P., Joulin, A., Bojanowski, P.: DINOv2: Learning Robust Visual Features...

    2024

  40. [40]

    The 2017 DAVIS Challenge on Video Object Segmentation

    Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Van Gool, L.: The 2017 DAVIS Challenge on Video Object Segmentation. arXiv preprint arXiv:1704.00675 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  41. [41]

    In: CVPR

    Qin, M., Li, W., Zhou, J., Wang, H., Pfister, H.: LangSplat: 3D Language Gaussian Splatting. In: CVPR. pp. 20051–20060 (2024)

    2024

  42. [42]

    In: ICML

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning Transfer- able Visual Models from Natural Language Supervision. In: ICML. pp. 8748–8763 (2021)

    2021

  43. [43]

    In: ICLR (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICLR (2021)

    2021

  44. [44]

    In: ICCV (2021)

    Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision Transformers for Dense Prediction. In: ICCV (2021)

    2021

  45. [45]

    NeurIPS37, 9153–9177 (2024) 18 Ding et al

    Shin, H., Kim, C., Hong, S., Cho, S., Arnab, A., Seo, P.H., Kim, S.: Towards Open-Vocabulary Semantic Segmentation Without Semantic Labels. NeurIPS37, 9153–9177 (2024) 18 Ding et al

    2024

  46. [46]

    In: ECCV

    Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor Segmentation and Support Inference from RGBD Images. In: ECCV. pp. 746–760 (2012)

    2012

  47. [47]

    DINOv3

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., Massa, F., Haziza, D., Wehrstedt, L., Wang, J., Darcet, T., Moutakanni, T., Sentana, L., Roberts, C., Vedaldi, A., Tolan, J., Brandt, J., Couprie, C., Mairal, J., Jégou, H., Labatut, P., Bojanowski, P.: DINOv3. arXiv preprint a...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  48. [48]

    Neurocomputing568, 127063 (2024)

    Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., Liu, Y.: RoFormer: Enhanced Trans- former with Rotary Position Embedding. Neurocomputing568, 127063 (2024)

    2024

  49. [49]

    In: ICML (2024)

    Sun, C., Yuan, Z., Xu, K., Mai, L., N, S., Chen, S., Marina, M.K.: Learning High- Frequency Functions Made Easy with Sinusoidal Positional Encoding. In: ICML (2024)

    2024

  50. [50]

    In: ECCV

    Suri, S., Walmer, M., Gupta, K., Shrivastava, A.: LiFT: A Surprisingly Simple Lightweight Feature Transform for Dense ViT Descriptors. In: ECCV. pp. 110–128 (2024)

    2024

  51. [51]

    SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features

    Tschannen, M., Gritsenko, A., Wang, X., Naeem, M.F., Alabdulmohsin, I., Parthasarathy, N., Evans, T., Beyer, L., Xia, Y., Mustafa, B., et al.: SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features. arXiv preprint arXiv:2502.14786 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  52. [52]

    In: ICCV

    Wang, J., Chen, K., Xu, R., Liu, Z., Loy, C.C., Lin, D.: CARAFE: Content-Aware ReAssembly of FEatures. In: ICCV. pp. 3007–3016 (2019)

    2019

  53. [53]

    In: CVPR (2018)

    Wang, X., Yu, K., Dong, C., Loy, C.C.: Recovering Realistic Texture in Image Super-Resolution by Deep Spatial Feature Transform. In: CVPR (2018)

    2018

  54. [54]

    In: ICLR (2026)

    Wimmer, T., Truong, P., Rakotosaona, M.J., Oechsle, M., Tombari, F., Schiele, B., Lenssen, J.E.: AnyUp: Universal Feature Upsampling. In: ICLR (2026)

    2026

  55. [55]

    In: ECCV

    Wu, Y., He, K.: Group Normalization. In: ECCV. pp. 3–19 (2018)

    2018

  56. [56]

    In: ECCV

    Wysoczańska, M., Siméoni, O., Ramamonjisoa, M., Bursuc, A., Trzciński, T., Pérez,P.:CLIP-DINOiser:TeachingCLIPafewDINOTricksforOpen-Vocabulary Semantic Segmentation. In: ECCV. pp. 320–337 (2024)

    2024

  57. [57]

    In: ICCV

    Xie, X., Lessen, J.E., Pons-Moll, G.: MVGBench: A Comprehensive Benchmark for Multi-view Generation Models. In: ICCV. pp. 8207–8218 (2025)

    2025

  58. [58]

    Demystifying CLIP Data

    Xu, H., Xie, S., Tan, X.E., Huang, P.Y., Howes, R., Sharma, V., Li, S.W., Ghosh, G., Zettlemoyer, L., Feichtenhofer, C.: Demystifying CLIP Data. arXiv preprint arXiv:2309.16671 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  59. [59]

    In: CVPR

    Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth Anything: Un- leashing the Power of Large-Scale Unlabeled Data. In: CVPR. pp. 10371–10381 (2024)

    2024

  60. [60]

    NeurIPS37, 21875–21911 (2024)

    Yang,L.,Kang,B.,Huang,Z.,Zhao,Z.,Xu,X.,Feng,J.,Zhao,H.:DepthAnything V2. NeurIPS37, 21875–21911 (2024)

    2024

  61. [61]

    In: ICCV

    Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid Loss for Language Image Pre-Training. In: ICCV. pp. 11941–11952 (2023)

    2023

  62. [62]

    a photo of a {label name}

    Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic Understanding of Scenes through the ADE20K Dataset. IJCV127(3), 302–321 (2019) RaysUp: Ultra-light Universal Feature Upsampling 19 RaysUp: Ultra-light Universal Feature Upsampling via Geometry-Aware Ray Representation — Supplementary Material — A Motivation of Spatial...

    2019

  63. [63]

    5 demonstrated consistent performance improvements, further validating the effectiveness and scalability of the proposed model

    (18) With approximately 4 hours of training, the results in Tab. 5 demonstrated consistent performance improvements, further validating the effectiveness and scalability of the proposed model. D.3 Upsampling from Any to Any Resolution. Following our semantic segmentation configurations, we evaluated the general- ization capability of each model for upsamp...