arxiv: 2604.18747 · v1 · submitted 2026-04-20 · 💻 cs.CV

Recognition: unknown

URoPE: Universal Relative Position Embedding across Geometric Spaces

Chensheng Peng, Depu Meng, Masayoshi Tomizuka, Quentin Herau, Wei Zhan, Yichen Xie, Yihan Hu

Pith reviewed 2026-05-10 04:13 UTC · model grok-4.3

classification 💻 cs.CV

keywords relative position embeddingrotary position embeddingtransformergeometric reasoningcross-viewcomputer vision3D projectionnovel view synthesis

0 comments

The pith

URoPE extends rotary position embeddings to cross-view and cross-dimensional geometry by sampling and projecting 3D ray points.

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

Standard rotary position embeddings work only on fixed 1D sequences or regular 2D/3D grids, which limits transformers in computer vision tasks that must reason across different camera views or between 2D images and 3D space. URoPE addresses this by sampling a few 3D points along each key-value camera ray at fixed depth anchors, projecting those points into the query image plane, and then applying ordinary 2D RoPE to the resulting coordinates. The method stays parameter-free, respects camera intrinsics, and does not depend on any shared global coordinate frame. It plugs directly into existing optimized RoPE attention kernels. Experiments on novel view synthesis, 3D detection, tracking, and depth estimation show consistent gains when URoPE replaces conventional position encodings.

Core claim

URoPE samples 3D points along the camera ray corresponding to each key or value patch at a small number of predefined depth anchors, projects these points into the query image plane using the query camera's intrinsics, and feeds the resulting 2D pixel coordinates into standard rotary position embedding. This produces a relative positional signal that is invariant to global coordinate choice and fully compatible with existing RoPE-accelerated attention implementations.

What carries the argument

Projection of sampled 3D points from key camera rays onto the query image plane, supplying 2D coordinates for standard RoPE.

If this is right

Transformers can apply relative positional encoding to attention between patches from different cameras or different dimensional representations without architectural changes.
Existing fast RoPE kernels remain usable, preserving training and inference speed.
Performance improves on novel view synthesis, 3D object detection, multi-object tracking, and monocular depth estimation.
The same embedding works for 2D-2D, 2D-3D, and temporal cross-frame relationships.

Where Pith is reading between the lines

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

The ray-sampling idea could extend to other geometric domains, such as spherical or non-Euclidean projections, if analogous sampling and projection rules are defined.
Multi-camera systems might reduce reliance on explicit extrinsic calibration inside the network once relative geometry is handled at the embedding level.
Replacing fixed depth anchors with learned or scene-adaptive sampling could further improve results on scenes with widely varying depth ranges.

Load-bearing premise

Sampling points at only a few fixed depth anchors along each ray and projecting them is sufficient to encode the relative geometry needed for cross-view and cross-dimensional reasoning.

What would settle it

Replacing standard RoPE with URoPE in the same transformer models on cross-view tasks and observing no accuracy gain or a clear drop in performance would show the projection step does not supply useful relative information.

Figures

Figures reproduced from arXiv: 2604.18747 by Chensheng Peng, Depu Meng, Masayoshi Tomizuka, Quentin Herau, Wei Zhan, Yichen Xie, Yihan Hu.

**Figure 1.** Figure 1: URoPE for Relative Position Embedding across Geometric Spaces. Previous relative position embeddings can only handle a shared geometric space (top), while URoPE focuses on the relative position embedding across geometric spaces (bottom). As a standard solution, position embeddings inject positional information into the permutation invariant Transformer architecture. Compared with absolute position embeddi… view at source ↗

**Figure 2.** Figure 2: Overview of URoPE. Left: in the same-view case, URoPE reduces to standard 2D RoPE. Middle: for cross-view interaction, a key-view pixel is lifted along its camera ray to multiple depth anchors and projected into the query view, producing headspecific positions along epipolar lines. Right: projected positions are grouped by depth anchors and assigned to different attention heads, enabling depth-anchored mu… view at source ↗

**Figure 3.** Figure 3: Qualitative Results for Novel View Synthesis. URoPE exploits relevant local information across views to synthesize sharper details. for the interaction between 3D queries and image features. As mentioned at the end of Sec. 3.2, URoPE can be generalized to this cross-dimensional interaction. Setup. To show the compatibility of URoPE, we consider both single-frame and long-horizontal multi-frame algorithms f… view at source ↗

**Figure 4.** Figure 4: Qualitative Comparison on 3D Object Detection. URoPE can help to identify small details from multiframe and multiview images. ther demonstrates that URoPE enhances the robust and coherent recognition of objects along the entire trajectories at multiple frames. 4.3 Stereo Depth Estimation Task Stereo depth estimation exploits parallax to predict depth, which requires information transfer between two camera… view at source ↗

**Figure 5.** Figure 5: Robustness Analysis. Although the model is trained with fixed camera focal length and 2 context views, URoPE can maintain a robust performance in out-ofdistribution cases. Fig. 5b adopts different context view selection with Tab. 2, so the metrics are not aligned [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: Region-wise Anchor Depth Importance. In the middle layer, the importance of depth anchors is highly related to the actual depth, while this correlation is not significant in both the shallow and deep layers [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗

read the original abstract

Relative position embedding has become a standard mechanism for encoding positional information in Transformers. However, existing formulations are typically limited to a fixed geometric space, namely 1D sequences or regular 2D/3D grids, which restricts their applicability to many computer vision tasks that require geometric reasoning across camera views or between 2D and 3D spaces. To address this limitation, we propose URoPE, a universal extension of Rotary Position Embedding (RoPE) to cross-view or cross-dimensional geometric spaces. For each key/value image patch, URoPE samples 3D points along the corresponding camera ray at predefined depth anchors and projects them into the query image plane. Standard 2D RoPE can then be applied using the projected pixel coordinates. URoPE is a parameter-free and intrinsics-aware relative position embedding that is invariant to the choice of global coordinate systems, while remaining fully compatible with existing RoPE-optimized attention kernels. We evaluate URoPE as a plug-in positional encoding for transformer architectures across a diverse set of tasks, including novel view synthesis, 3D object detection, object tracking, and depth estimation, covering 2D-2D, 2D-3D, and temporal scenarios. Experiments show that URoPE consistently improves the performance of transformer-based models across all tasks, demonstrating its effectiveness and generality for geometric reasoning. Our project website is: https://urope-pe.github.io/.

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.

Desk Editor's Note private letter to a colleague

URoPE extends RoPE to cross-view geometry via ray sampling and projection, but fixed depth anchors look like the weakest link for continuous 3D relations.

read the letter

The core move here is sampling a handful of 3D points along each key ray at fixed depth anchors, projecting those points into the query image plane, and then feeding the resulting 2D coordinates into ordinary RoPE. That single step lets the same rotary mechanism handle 2D-2D, 2D-3D, and temporal cases without changing the attention kernel or adding parameters. It is also invariant to the choice of world coordinate frame because everything is expressed relative to the query camera. Those are the two things worth knowing up front: the construction is parameter-free and kernel-compatible, and it is derived straight from standard camera geometry rather than learned from scratch. The paper shows this plug-in on novel-view synthesis, 3D detection, tracking, and depth estimation, and the abstract claims gains across all of them. That breadth is useful; most positional-encoding papers stay inside one narrow task. The math itself is straightforward once the projection is written down, and the lack of fitted parameters removes one common source of overfitting. The soft spot is exactly the one the stress-test flags. Fixed, non-adaptive depth anchors mean the sampled points only approximate the true ray intersection when scene depths happen to lie near those anchors. Large baselines or depth ranges outside the chosen set will produce projected coordinates that no longer match the actual 3D displacement, so the rotary angles stop being faithful geometric encodings. If the full paper contains ablations that vary the number and placement of anchors and still show stable gains, that would address the concern; the abstract does not. The experiments are described only at the level of “consistent improvements,” with no numbers, baselines, or error breakdowns visible here, so the empirical case remains thin until the tables are checked. This is the kind of paper a reading group on geometric transformers or multi-view attention would want to discuss, mainly to see whether the projection trick actually survives real parallax. I would send it to peer review because the idea is new, the derivation is clean, and the compatibility claim is testable; a referee can ask for the missing ablations and quantitative results without starting from zero.

Referee Report

2 major / 3 minor

Summary. The paper proposes URoPE, a universal extension of Rotary Position Embedding (RoPE) to cross-view and cross-dimensional geometric spaces. For each key/value patch, it samples 3D points along the camera ray at a small number of predefined depth anchors, projects them into the query image plane using camera intrinsics, and applies standard 2D RoPE to the resulting 2D coordinates. The method is presented as parameter-free, intrinsics-aware, invariant to global coordinate choice, and directly compatible with existing RoPE-optimized attention kernels. Experiments across novel view synthesis, 3D object detection, object tracking, and depth estimation (covering 2D-2D, 2D-3D, and temporal settings) report consistent performance gains when URoPE is used as a plug-in positional encoding in transformer models.

Significance. If the central construction holds, URoPE would provide a practical, parameter-free mechanism for injecting relative geometric information into attention without kernel modifications or learned parameters, extending RoPE beyond fixed 1D/2D/3D grids. This could improve geometric reasoning in multi-view and 3D CV tasks. The derivation from standard camera projection and existing RoPE is a strength, as is the explicit invariance and kernel compatibility.

major comments (2)

[Method section (URoPE construction)] Method section (URoPE construction): the central step samples a small number of 3D points at fixed, non-adaptive depth anchors along each key ray and projects them into the query plane before applying 2D RoPE. This discrete approximation is load-bearing for the universality claim; it is unclear whether the projected coordinates remain faithful proxies for relative 3D displacement when scene depths lie between or outside the anchors or when parallax is large, as the skeptic note highlights.
[Experiments section] Experiments section: while consistent improvements are claimed across four tasks, the provided text supplies no quantitative tables, ablation results on anchor count or placement, baseline details, or error analysis conditioned on depth or baseline variation. This leaves the empirical support for the geometric-sufficiency assumption only partially verifiable.

minor comments (3)

[Abstract and method] Abstract and method: the number and selection criterion for the 'predefined depth anchors' are not stated, which affects reproducibility and should be specified (e.g., uniform in disparity or log-depth).
[Notation] Notation: the exact projection equation mapping the sampled 3D points to query-plane pixel coordinates should be written explicitly, including the role of intrinsics matrices.
[Figures] Figures: any diagram illustrating the ray sampling and projection step would benefit from explicit depth-anchor labels and an example of a large-baseline case.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment of URoPE's core construction and its potential utility. We address each major comment below with clarifications and commit to targeted revisions that strengthen the manuscript without altering the method.

read point-by-point responses

Referee: [Method section (URoPE construction)] Method section (URoPE construction): the central step samples a small number of 3D points at fixed, non-adaptive depth anchors along each key ray and projects them into the query plane before applying 2D RoPE. This discrete approximation is load-bearing for the universality claim; it is unclear whether the projected coordinates remain faithful proxies for relative 3D displacement when scene depths lie between or outside the anchors or when parallax is large, as the skeptic note highlights.

Authors: The discrete sampling is indeed an approximation, but it is designed to produce view-dependent 2D relative coordinates in the query plane rather than to reconstruct exact 3D displacements. Because the subsequent 2D RoPE operates directly on these projected pixel locations, the attention mechanism receives a geometrically consistent relative signal that is invariant to global coordinate choice and depends only on camera intrinsics and the relative pose between views. Logarithmically spaced anchors (typically 4–8) are chosen to span the depth range of each dataset; intermediate depths produce interpolated projections that remain useful for attention. Large parallax is handled naturally because the projection is performed per query-key pair. We will expand the method section with a formal justification of the approximation, explicit anchor-selection guidelines, and a short derivation showing that the projected coordinates encode the essential relative geometry for cross-view attention. We will also add a sensitivity analysis on anchor count and spacing. revision: partial
Referee: [Experiments section] Experiments section: while consistent improvements are claimed across four tasks, the provided text supplies no quantitative tables, ablation results on anchor count or placement, baseline details, or error analysis conditioned on depth or baseline variation. This leaves the empirical support for the geometric-sufficiency assumption only partially verifiable.

Authors: The reviewed manuscript version omitted the full result tables and ablations that appear in the complete draft. We will insert comprehensive quantitative tables reporting absolute metrics (PSNR/SSIM for NVS, mAP for detection, MOTA for tracking, RMSE for depth) together with relative gains over standard RoPE, learned positional embeddings, and task-specific baselines. New ablations will quantify the effect of anchor count (4, 8, 16) and placement (linear vs. log-spaced) across all four tasks. We will also add error-analysis figures and tables stratified by depth bins and baseline distance to directly address the geometric-sufficiency assumption. These additions will make the empirical claims fully verifiable. revision: yes

Circularity Check

0 steps flagged

No significant circularity; URoPE is a direct geometric construction from standard RoPE and camera projection

full rationale

The paper defines URoPE explicitly as sampling a fixed set of 3D points at predefined depth anchors along each key ray, projecting those points into the query image plane via standard camera intrinsics, and then feeding the resulting 2D coordinates into unmodified 2D RoPE. This construction is parameter-free, follows directly from projective geometry and the existing RoPE equations, and contains no fitted parameters, self-referential equations, or predictions that reduce to prior fits. No load-bearing self-citations, uniqueness theorems, or ansatzes imported from prior author work are invoked for the core claim. Experimental gains are reported as independent empirical outcomes rather than tautological consequences of the definition. The discrete-anchor choice is an explicit modeling decision whose sufficiency is tested externally, not assumed by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

URoPE relies on standard assumptions from computer vision geometry and the existing RoPE formulation. No new free parameters are introduced as it is explicitly parameter-free. No new entities are postulated.

axioms (2)

domain assumption Known camera intrinsics allow accurate projection of 3D points to 2D image planes
The projection step depends on this standard assumption in multi-view geometry.
standard math Rotary Position Embedding can be directly applied to the projected 2D coordinates
Builds directly on the RoPE mechanism without modification.

pith-pipeline@v0.9.0 · 5575 in / 1498 out tokens · 45065 ms · 2026-05-10T04:13:18.206709+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

44 extracted references · 14 canonical work pages · 8 internal anchors

[1]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11621–11631 (2020) 3, 9, 21

2020
[2]

In: European conference on computer vision

Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End- to-end object detection with transformers. In: European conference on computer vision. pp. 213–229. Springer (2020) 1, 3

2020
[3]

In: Pro- ceedingsofthe57thannualmeetingoftheassociationforcomputationallinguistics

Dai, Z., Yang, Z., Yang, Y., Carbonell, J.G., Le, Q., Salakhutdinov, R.: Transformer-xl: Attentive language models beyond a fixed-length context. In: Pro- ceedingsofthe57thannualmeetingoftheassociationforcomputationallinguistics. pp. 2978–2988 (2019) 2

2019
[4]

In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 3, 4

Dao,T.,Fu,D.Y.,Ermon,S.,Rudra,A.,Ré,C.:FlashAttention:Fastandmemory- efficient exact attention with IO-awareness. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 3, 4

2022
[5]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Deitke, M., Schwenk, D., Salvador, J., Weihs, L., Michel, O., VanderBilt, E., Schmidt, L., Ehsani, K., Kembhavi, A., Farhadi, A.: Objaverse: A universe of annotated 3d objects. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 13142–13153 (2023) 3, 8, 21

2023
[6]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 3

work page internal anchor Pith review Pith/arXiv arXiv 2010
[7]

In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

Du, Y., Smith, C., Tewari, A., Sitzmann, V.: Learning to render novel views from wide-baseline stereo pairs. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 4970–4980 (2023) 4

2023
[8]

The Llama 3 Herd of Models

Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Let- man, A., Mathur, A., Schelten, A., Vaughan, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024) 3

work page internal anchor Pith review Pith/arXiv arXiv 2024
[9]

He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016) 21 16 Y. Xie et al

2016
[10]

In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition

He, Y., Yan, R., Fragkiadaki, K., Yu, S.I.: Epipolar transformers. In: Proceedings of the ieee/cvf conference on computer vision and pattern recognition. pp. 7779–7788 (2020) 4

2020
[11]

Heo, B., Park, S., Han, D., Yun, S.: Rotary position embedding for vision trans- former.In:EuropeanConferenceonComputerVision.pp.289–305.Springer(2024) 3, 21

2024
[12]

Lrm: Large reconstruction model for single image to 3d,

Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400 (2023) 4

work page arXiv 2023
[13]

Lvsm: A large view synthesis model with minimal 3d inductive bias.arXiv preprint arXiv:2410.17242, 2024

Jin, H., Jiang, H., Tan, H., Zhang, K., Bi, S., Zhang, T., Luan, F., Snavely, N., Xu, Z.: Lvsm: A large view synthesis model with minimal 3d inductive bias. arXiv preprint arXiv:2410.17242 (2024) 1, 4, 5, 8, 9, 11, 13, 14, 19

work page arXiv 2024
[14]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Li, J., Wang, P., Xiong, P., Cai, T., Yan, Z., Yang, L., Liu, J., Fan, H., Liu, S.: Practical stereo matching via cascaded recurrent network with adaptive correla- tion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16263–16272 (2022) 1

2022
[15]

arXiv preprint arXiv:2507.10496 (2025) 2, 4, 8, 9, 11, 13, 19, 21

Li, R., Yi, B., Liu, J., Gao, H., Ma, Y., Kanazawa, A.: Cameras as relative po- sitional encoding. arXiv preprint arXiv:2507.10496 (2025) 2, 4, 8, 9, 11, 13, 19, 21

work page arXiv 2025
[16]

IEEE Transactions on Pattern Analysis and Machine Intelligence47(3), 2020–2036 (2024) 4

Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Yu, Q., Dai, J.: Bevformer: learningbird’s-eye-viewrepresentationfromlidar-cameraviaspatiotemporaltrans- formers. IEEE Transactions on Pattern Analysis and Machine Intelligence47(3), 2020–2036 (2024) 4

2020
[17]

Depth Anything 3: Recovering the Visual Space from Any Views

Lin, H., Chen, S., Liew, J., Chen, D.Y., Li, Z., Shi, G., Feng, J., Kang, B.: Depth anything 3: Recovering the visual space from any views. arXiv preprint arXiv:2511.10647 (2025) 21

work page internal anchor Pith review arXiv 2025
[18]

In: Proceedings of the IEEE/CVF inter- national conference on computer vision

Liu, R., Wu, R., Van Hoorick, B., Tokmakov, P., Zakharov, S., Vondrick, C.: Zero- 1-to-3: Zero-shot one image to 3d object. In: Proceedings of the IEEE/CVF inter- national conference on computer vision. pp. 9298–9309 (2023) 21

2023
[19]

In: European conference on computer vision

Liu, Y., Wang, T., Zhang, X., Sun, J.: Petr: Position embedding transformation for multi-view 3d object detection. In: European conference on computer vision. pp. 531–548. Springer (2022) 1, 4, 9, 10, 21

2022
[20]

In: Proceedings of the IEEE/CVF in- ternational conference on computer vision

Meng, D., Chen, X., Fan, Z., Zeng, G., Li, H., Yuan, Y., Sun, L., Wang, J.: Con- ditional detr for fast training convergence. In: Proceedings of the IEEE/CVF in- ternational conference on computer vision. pp. 3651–3660 (2021) 1

2021
[21]

arXiv preprint arXiv:2310.10375 (2023) 2, 4

Miyato, T., Jaeger, B., Welling, M., Geiger, A.: Gta: A geometry-aware attention mechanism for multi-view transformers. arXiv preprint arXiv:2310.10375 (2023) 2, 4

work page arXiv 2023
[22]

Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation

Press, O., Smith, N.A., Lewis, M.: Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409 (2021) 2

work page internal anchor Pith review arXiv 2021
[23]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

Sajjadi, M.S., Meyer, H., Pot, E., Bergmann, U., Greff, K., Radwan, N., Vora, S., Lučić, M., Duckworth, D., Dosovitskiy, A., et al.: Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 6229–6238 (2022) 1

2022
[24]

In: Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers)

Shaw, P., Uszkoreit, J., Vaswani, A.: Self-attention with relative position represen- tations. In: Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers). pp. 464–468 (2018) 2 URoPE17

2018
[25]

Advances in Neural Information Processing Systems34, 19313–19325 (2021) 2, 4

Sitzmann, V., Rezchikov, S., Freeman, B., Tenenbaum, J., Durand, F.: Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems34, 19313–19325 (2021) 2, 4

2021
[26]

In: 2012 IEEE/RSJ international conference on intelligent robots and systems

Sturm, J., Engelhard, N., Endres, F., Burgard, W., Cremers, D.: A benchmark for the evaluation of rgb-d slam systems. In: 2012 IEEE/RSJ international conference on intelligent robots and systems. pp. 573–580. IEEE (2012) 3, 10, 11

2012
[27]

Neurocomputing568, 127063 (2024) 2, 3, 7

Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., Liu, Y.: Roformer: Enhanced trans- former with rotary position embedding. Neurocomputing568, 127063 (2024) 2, 3, 7

2024
[28]

In: DAGM German Conference on Pattern Recognition

Tang, Y., Dorn, S., Savani, C.: Center3d: Center-based monocular 3d object detec- tion with joint depth understanding. In: DAGM German Conference on Pattern Recognition. pp. 289–302. Springer (2020) 14

2020
[29]

LLaMA: Open and Efficient Foundation Language Models

Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) 3

work page internal anchor Pith review Pith/arXiv arXiv 2023
[30]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Ummenhofer, B., Zhou, H., Uhrig, J., Mayer, N., Ilg, E., Dosovitskiy, A., Brox, T.: Demon: Depth and motion network for learning monocular stereo. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 5038–5047 (2017) 3, 10, 11

2017
[31]

Advances in neural information pro- cessing systems30(2017) 1, 2, 3

Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017) 1, 2, 3

2017
[32]

In: Proceedings of the Computer Vision and Pattern Recognition Conference

Wang, J., Chen, M., Karaev, N., Vedaldi, A., Rupprecht, C., Novotny, D.: Vggt: Visual geometry grounded transformer. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 5294–5306 (2025) 21

2025
[33]

Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., et al.: Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191 (2024) 3

work page internal anchor Pith review Pith/arXiv arXiv 2024
[34]

In: Proceedings of the IEEE/CVF international conference on computer vision

Wang, S., Liu, Y., Wang, T., Li, Y., Zhang, X.: Exploring object-centric tempo- ral modeling for efficient multi-view 3d object detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 3621–3631 (2023) 1, 4, 9, 10, 21

2023
[35]

In: Conference on robot learning

Wang, Y., Guizilini, V.C., Zhang, T., Wang, Y., Zhao, H., Solomon, J.: Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In: Conference on robot learning. pp. 180–191. PMLR (2022) 4

2022
[36]

arXiv preprint arXiv:2601.15275 (2026) 4, 8, 9, 21

Wu, Y., Jeon, M., Chang, J.H.R., Tuzel, O., Tulsiani, S.: Rayrope: Projective ray positional encoding for multi-view attention. arXiv preprint arXiv:2601.15275 (2026) 4, 8, 9, 21

work page arXiv 2026
[37]

In: Proceedings of the IEEE international conference on computer vision

Xiao, J., Owens, A., Torralba, A.: Sun3d: A database of big spaces reconstructed using sfm and object labels. In: Proceedings of the IEEE international conference on computer vision. pp. 1625–1632 (2013) 3, 10, 11

2013
[38]

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2026) 2, 4, 5

Xie, Y., Peng, C., Abdelfattah, M., Hu, Y., Yang, J., Higgins, E., Brigden, R., Tomizuka, M., Zhan, W.: Raynova: Scale-temporal autoregressive world modeling in ray space. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2026) 2, 4, 5

2026
[39]

arXiv preprint arXiv:2411.01123 (2024) 4

Xie, Y., Xu, C., Peng, C., Zhao, S., Ho, N., Pham, A.T., Ding, M., Tomizuka, M., Zhan, W.: X-drive: Cross-modality consistent multi-sensor data synthesis for driving scenarios. arXiv preprint arXiv:2411.01123 (2024) 4

work page arXiv 2024
[40]

Xie et al

Xie, Y., Xu, C., Rakotosaona, M.J., Rim, P., Tombari, F., Keutzer, K., Tomizuka, M., Zhan, W.: Sparsefusion: Fusing multi-modal sparse representations for multi- 18 Y. Xie et al. sensor 3d object detection. In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision. pp. 17591–17602 (2023) 4

2023
[41]

IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 1, 4, 10, 11, 21

Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Yu, F., Tao, D., Geiger, A.: Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 1, 4, 10, 11, 21

2023
[42]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Yin, T., Zhou, X., Krahenbuhl, P.: Center-based 3d object detection and track- ing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11784–11793 (2021) 9

2021
[43]

Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo magnification: Learn- ingviewsynthesisusingmultiplaneimages.arXivpreprintarXiv:1805.09817(2018) 3, 8, 13, 21

work page internal anchor Pith review arXiv 2018
[44]

Deformable DETR: Deformable Transformers for End-to-End Object Detection

Zhu, X., Su, W., Lu, L., Li, B., Wang, X., Dai, J.: Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159 (2020) 4 URoPE19 In this appendix, we exhibit additional experiment results in Sec. A. After- wards, the implementation details of our experiments are explained in Sec. B. Finally, we discuss some li...

work page internal anchor Pith review arXiv 2010