D$^2$-World: An Efficient World Model through Decoupled Dynamic Flow

Bingbing Liu; Haiming Zhang; Shuguang Cui; Xu Yan; Ying Xue; Zhen Li; Zixuan Guo

REVIEW 3 major objections 6 minor 14 references

D$^2$-World: An Efficient World Model through Decoupled Dynamic Flow

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper claims that point-cloud forecasting for autonomous driving is best reformulated as occupancy prediction followed by a decoupled dynamic flow, warping only dynamic voxels while moving static voxels by ego pose, which reaches a…

desk verdict Solid workshop solution with a genuinely useful decoupling trick; the matched-split gains are large, but the loss-metric mismatch and lack of error bars are the real weaknesses. read the letter →

arxiv 2411.17027 v1 pith:B553SPQW submitted 2024-11-26 cs.CV

Haiming Zhang , Xu Yan , Ying Xue , Zixuan Guo , Shuguang Cui , Zhen Li , Bingbing Liu This is my paper · ORCID

classification cs.CV

keywords worldmodelpointcloudforecasting4Doccupancydecoupleddynamicflownon-autoregressivepredictionautonomousdrivingChamferDistance

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

This paper argues that visual point-cloud forecasting for autonomous driving is best reformulated as two decoupled tasks: first predict semantic occupancy from past camera images, then forecast future occupancy in a single non-autoregressive pass. The key move is to separate static voxels, which merely follow the ego vehicle's pose changes, from dynamic voxels, which are warped by a learned voxel flow. That decoupling removes most of the prediction burden, since the static majority of the scene needs no learned motion model. On the OpenScene benchmark the resulting system, D2-World, reaches a Chamfer Distance of 0.79 with one model, ranking second, and trains in about 30% of the baseline's wall-clock time. The reason to care is that efficient, single-stage forecasting makes large-scale world-model training practical for end-to-end driving.

What carries the argument

The load-bearing object is the decoupled dynamic flow: the flow decoder predicts future flow only for dynamic voxels, while static voxels are transformed directly by future ego poses, and a light CNN refinement cleans the coarse warped features. It is carried by SALT (spatial-aware local-temporal) attention blocks, which replace the transformer's MLP and feed-forward network with 2D and 3D convolutions so the model keeps spatial structure while attending over time. The rendering loss is the same ray-wise cross-entropy used by ViDAR, applied to the predicted occupancy density to produce point clouds.

What would settle it

A direct check: during stage-II training, log the ray-wise cross-entropy and the validation Chamfer Distance every epoch on a fixed split. If the loss decreases while Chamfer Distance stalls or rises for several consecutive epochs, the proxy-loss assumption is false; if the two track together, the paper's supervision choice is validated.

Watch

Extended reading notes

Core claim

The central claim is that forecasting future point clouds can be decomposed into occupancy prediction plus a decoupled dynamic flow. Given past camera images, a first stage produces semantic 3D occupancy frame by frame. A second stage encodes those occupancies as spacetime tokens, runs a spatial-temporal transformer, and predicts voxel flow for every future timestamp in parallel, then renders the warped occupancy into point clouds with differentiable volume rendering. Static voxels are not predicted at all; they are moved by ego-pose transformation. This decomposition is what lets a single non-autoregressive model beat the autoregressive ViDAR baseline by a large margin and cut training cost to about one third.

Load-bearing premise

The load-bearing premise is that the rendering-based ray-wise cross-entropy loss used to train the flow is a sufficient stand-in for the Chamfer Distance used for evaluation, because the paper never shows that lowering one lowers the other.

Editorial extensions

If this is right

Training time drops from 23.50 GPU-hours to 7.14 GPU-hours, so full-dataset world-model training becomes feasible without the memory-heavy tricks the baseline needs.
Non-autoregressive decoding predicts all future timestamps at once, eliminating the sequential cost and potential error compounding of frame-by-frame autoregression.
Because static voxels are handled by pose, the learned flow concentrates its capacity on moving objects, where forecasting error is concentrated.
The two-stage decomposition lets each stage improve independently, and the paper's ablation shows that better occupancy prediction directly improves forecasting quality.
The same rendering-based supervision keeps the framework compatible with camera-only setups, since the point cloud is generated from predicted occupancy rather than raw sensor data.

Reading between the lines

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

If the decoupling is as effective as reported, a natural follow-up is to measure the residual error attributable to dynamic-object motion alone, which would isolate what the learned flow module must still capture.
The paper leaves implicit that the same occupancy-to-flow pipeline could be trained directly against a differentiable surrogate of Chamfer Distance, which would directly test whether the rendering loss is the right supervision.
The SALT block design, mixing convolutional spatial processing with 3D-convolutional temporal mixing, could transfer to other dense 4D prediction tasks such as video prediction or LiDAR simulation, since it does not depend on camera intrinsics.

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.

Desk Editor's Note

read the letter

Quick take: this is a solid workshop challenge report. The core idea—decouple dynamic and static voxels, warp dynamic ones with predicted flow and static ones with ego pose—is simple and effective. The matched comparison on the 1/8 mini split shows D2-World at 0.71 CD against ViDAR's 1.58, a big improvement, not the modest 0.06 the reader remembered; that 0.06 figure comes from the Online Server rows where the training splits differ (Mini vs Full), and those rows are indeed confounded. So the paper's central claim survives, but the leaderboard numbers should be read with that caveat.

What's genuinely useful: the non-autoregressive design trains ~3x faster (Table 2), and Table 3 gives a clean ablation showing semantic occupancy helps and GT occupancy only marginally improves results (0.71 to 0.69), which tells you the renderer/flow are already near the ceiling of this proxy task. The code release is a plus.

The soft spots are real but not disqualifying. The world model's only supervision is the ray-wise CE render loss borrowed from ViDAR, and the paper never shows this CE loss correlates with the CD metric. That's a legitimate gap in the write-up. But the method actually ranked second on the leaderboard, which uses CD, so empirically the surrogate does the job. Still, a direct flow evaluation (e.g., against scene flow) or a CD-based loss would strengthen the case. The absence of error bars is also a minor issue; these are single-run results typical of challenge reports.

This is not a paradigm shift, but it is a clean, useful combination of existing ideas. It deserves referee time as a workshop solution, and with a few additions (matched baselines on the full split, error bars, flow analysis) it could be a decent conference paper. I'd bring it to a reading group if the topic is 4D occupancy or world models.

Referee Report

3 major / 6 minor

Summary. The paper describes D2-World, the second-place solution to the CVPR 2024 Predictive World Model Challenge. The method is a two-stage pipeline: Stage I predicts semantic occupancy from multi-view images using a BEVDet-like network, and Stage II uses a non-autoregressive spatial-temporal transformer with SALT attention blocks to forecast future BEV occupancy. The key design choice is decoupled dynamic flow: voxels are split into dynamic and static groups using the predicted semantics, dynamic voxels are warped by a learned flow, static voxels are warped by ego pose, and the result is rendered into a point cloud with the ViDAR ray-wise cross-entropy loss. The paper reports a Chamfer Distance of 0.79 on the online server, second place on the leaderboard, and roughly 3x faster training than the ViDAR baseline.

Significance. If the reported results hold, the two-stage, non-autoregressive formulation is a useful efficiency improvement over the autoregressive ViDAR baseline, and the decoupled dynamic-flow idea is an interesting way to simplify occupancy forecasting. The paper also releases code and provides a detailed training-efficiency table, which are concrete assets. However, the current evidence is weaker than the claims suggest: the headline comparison in Table 1 mixes training splits, and the only supervision for the learned flow is a proxy loss whose alignment with the evaluation metric is not demonstrated. The contribution of the decoupled flow itself is therefore not yet isolated, which limits the strength of the central claim.

major comments (3)

[§3.2, Table 1] The headline comparison is confounded by unequal training splits. The row 'D2-World (ours) Full Online Server' (CD 0.79) is presented as surpassing 'ViDAR (baseline) Mini Online Server' (CD 1.59), but the ViDAR row was trained on 1/8 of the data while D2-World was trained on the full split. Under matched conditions the margin is much smaller: on the Mini Online Server split, D2-World vanilla achieves 1.53 vs. ViDAR's 1.59, and the only matched comparison with decoupled flow is on the Mini test split (0.71 vs. 1.58). The abstract and introduction should state exactly which comparison supports the 'large margin' claim, and the paper should report a matched ViDAR baseline trained on the same split as each D2-World row.
[§2.2 'Rendering & Losses', §3.1 'Metric'] The only supervision for the predicted dynamic flow is the ViDAR ray-wise cross-entropy loss applied to the rendered point cloud, yet no evidence is provided that minimizing this loss also minimizes the Chamfer Distance used for evaluation. The paper never measures the predicted flow against ground-truth scene flow, nor does it compare the full model against a static-only variant that warps all voxels by ego pose. Without such a comparison, the reported CD advantage cannot be attributed to the learned decoupled dynamic flow rather than to the static-warping path. A concrete remedy is to report flow end-point error on annotated scenes and to include an ablation that removes the dynamic-flow branch entirely.
[§3.2 'The Effects of Occupancy Performance', Table 3] The statement that 'the performance does not significantly improve even when ground truth occupancy ... is used' is used to argue that the rendering/world-model stage is the bottleneck. However, no statistical test, standard deviation, or multiple-seed results are reported, so 'not significantly' is not substantiated. Since this observation is load-bearing for the paper's interpretation of where the remaining error lies, please provide variance estimates over validation scenes or a paired significance test.

minor comments (6)

[Abstract and §1] The training-speed claims are inconsistent: the abstract says 'more than 300% faster' while the introduction says 'increased the training speed by 200%'. Table 2 shows D2-World total 7.14h vs. ViDAR 23.50h, which is about a 3.3x speedup. Please use one consistent, unambiguous formulation and define what 'x% faster' means relative to what baseline.
[Table 1] The table header contains the typo 'Poin cloud forecasting performance'; it should read 'Point cloud forecasting performance'.
[§2.2 'Spatial-Temporal Transformer'] The acronym SALT is introduced without a definition or a reference to a prior work that defines it; please give a full name or cite the source.
[§2.2 'Decoupled Dynamic Flow'] The phrase 'warping the existing observations through voxel flow' is vague about whether the flow is defined in 3D space or in the BEV plane, and whether it operates on voxel centers or features. A precise definition of the flow field and the warping operation would help reproducibility.
[Table 3] For Versions E and H, 'use GT' should specify whether the ground-truth occupancy is fed as input to the world model or used as supervision, and what the 100.0 mIoU/IoU columns mean for those rows. Also, the ViDAR row has dashes for mIoU and IoU; please state why these are not reported.
[References] Reference [10] (Press et al., 'Train short, test long') is about length extrapolation for attention and is not directly related to video prediction; if it is cited for SALT's temporal attention, please clarify the connection or add a more relevant citation.

Circularity Check

0 steps flagged · score 0.0 of 10

No material circularity: the method is trained with supervised losses on the dataset and evaluated on the external OpenScene benchmark, and the borrowed ViDAR rendering loss is a training-signal choice rather than a definitional equivalence.

full rationale

The paper's derivation chain is not circular. Stage I trains an occupancy network with class-weighted cross-entropy and Lovasz losses against occupancy labels, and Stage II trains a world model with the ViDAR ray-wise cross-entropy rendering loss against rendered point clouds; neither training signal is constructed from the evaluation metric (Chamfer Distance), and the model is evaluated on held-out data via an online server. The statement 'We utilize the same rendering process and losses as ViDAR [13] for optimizing the point cloud forecasting' borrows a supervision scheme from the baseline, but this is not equivalent to predicting the evaluation metric by construction. The decoupled dynamic flow is ablated in Table 1 (D2-World vanilla vs D2-World), and Table 3 varies occupancy quality, so the contributions are empirically tested rather than assumed. The paper does cite prior work by overlapping authors ([12], [14]) for the occupancy head and as examples of occupancy networks, but these citations are not load-bearing in the sense of importing an unverified uniqueness theorem or defining the target result in terms of itself; the cited components are replaceable and are evaluated through ablations and benchmark results. The legitimate weakness is that the ray-wise cross-entropy loss may not be perfectly aligned with Chamfer Distance, and the predicted flow itself is not directly compared to ground-truth scene flow. That is a correctness and validation concern, not circularity: an unverified proxy loss is still an external training signal, not a re-statement of the evaluation metric. No equation in the paper reduces a predicted quantity to a fitted input, and no claim rests solely on a self-citation chain. Therefore the appropriate circularity score is 0.

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

The central claim rests on standard modeling assumptions and hand-chosen hyperparameters. There are no new physical entities. The unstated patch size and reliance on ViDAR rendering are the most significant unaccounted choices.

free parameters (4)

Occupancy grid resolution = 200x200x16
Defines the 3D scene discretization for both stages; chosen by hand to balance detail and compute.
Image input resolution = 512x1408
Used for stage I image encoding; affects occupancy quality and training cost.
Learning rates and batch sizes = Stage I: 2e-4, batch 24; Stage II: 1e-3, batch 16 (ablation batch 8)
Training hyperparameters selected by hand for each stage; no search or derivation is reported.
Voxel patch size P
Used in Section 2.2 to convert BEV embeddings into 2D patches; the paper does not report its value, making exact re-implementation ambiguous.

assumptions (4)

domain assumption Voxel flow plus ego-pose transformation can represent the future occupancy of a scene.
The method assumes dynamic voxels can be warped by a flow field and static voxels by ego pose alone. This ignores object deformation and complex interactions.
domain assumption The ray-wise cross-entropy rendering loss from ViDAR is a sufficient training signal for point cloud forecasting.
Section 2.2 adopts the ViDAR rendering and losses without demonstrating that optimizing this proxy also minimizes the evaluation metric, Chamfer Distance.
domain assumption Semantic occupancy from stage I can reliably separate dynamic and static voxels.
The decoupled flow relies on accurate semantics; if the occupancy network misclassifies motion, the warping step will be wrong.
domain assumption OpenScene occupancy labels align with LiDAR point clouds after grid sampling.
Section 3.1 states that grid sampling aligns annotations from [-50m, -50m, -4m, 50m, 50m, 4m] to the LiDAR range; any misalignment would propagate to the rendered point clouds.

how reviews work

0 comments

Cite this review

Pith. "Pith review of D$^2$-World: An Efficient World Model through Decoupled Dynamic Flow." pith.science (2026). https://pith.science/paper/B553SPQW

@misc{pith2026241117027,
  author       = {Pith},
  title        = {Pith review of: D$^2$-World: An Efficient World Model through Decoupled Dynamic Flow},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B553SPQW}},
  note         = {Machine review of arXiv:2411.17027}
}

abstract

This technical report summarizes the second-place solution for the Predictive World Model Challenge held at the CVPR-2024 Workshop on Foundation Models for Autonomous Systems. We introduce D$^2$-World, a novel World model that effectively forecasts future point clouds through Decoupled Dynamic flow. Specifically, the past semantic occupancies are obtained via existing occupancy networks (e.g., BEVDet). Following this, the occupancy results serve as the input for a single-stage world model, generating future occupancy in a non-autoregressive manner. To further simplify the task, dynamic voxel decoupling is performed in the world model. The model generates future dynamic voxels by warping the existing observations through voxel flow, while remaining static voxels can be easily obtained through pose transformation. As a result, our approach achieves state-of-the-art performance on the OpenScene Predictive World Model benchmark, securing second place, and trains more than 300% faster than the baseline model. Code is available at https://github.com/zhanghm1995/D2-World.

Figures

Figures reproduced from arXiv: 2411.17027 by the authors.

**Figure 1.** The overall pipeline of D2 -World. In the first stage, we train a single-frame occupancy network, and in the second stage, we train a world model that takes past occupancy as input, forecasting future point clouds. ages with T timestamps, the first stage predicts occupancy frame-by-frame, aiming to recover a rich 3D dense representation from the 2D images. In the second stage, we approach this as a 4D point cloud … view at source ↗

**Figure 2.** Inner structure of SALT & warping and refinement. (a) The detailed structures of SALT, which replace the MLP and FFN (Feed Forward Network) in vanilla transformer with 2D convolutions and 3D convolutions respectively for capturing spatialtemporal dependencies. (b) We decouple the flow with the dynamic and static flow and warp the feature of the current frame for forecasting the future frame. The refinement module… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 6 canonical work pages

[1]

nuplan: A closed-loop ml-based plan- ning benchmark for autonomous vehicles

Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based plan- ning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 3

arXiv 2021
[2]

Openscene: The largest up-to- date 3d occupancy prediction benchmark in autonomous driving

OpenScene Contributors. Openscene: The largest up-to- date 3d occupancy prediction benchmark in autonomous driving. https://github.com/OpenDriveLab/ OpenScene, 2023. 1, 3

work page 2023
[3]

Faceformer: Speech-driven 3d facial anima- tion with transformers

Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial anima- tion with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 18770–18780, 2022. 3

work page 2022
[4]

Bevdet: High-performance multi-camera 3d object de- tection in bird-eye-view

Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object de- tection in bird-eye-view. arXiv preprint arXiv:2112.11790,

arXiv
[5]

Point Cloud Forecasting as a Proxy for 4D Occu- pancy Forecasting

Tarasha Khurana, Peiyun Hu, David Held, and Deva Ra- manan. Point Cloud Forecasting as a Proxy for 4D Occu- pancy Forecasting. In CVPR, 2023. 3

work page 2023
[6]

Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo

Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo. arXiv preprint arXiv:2209.10248, 2022. 2

arXiv 2022
[7]

Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers

Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chong- hao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers. In European con- ference on computer vision, pages 1–18. Springer, 2022. 1

2022
[8]

Swin transformer: Hierarchical vision transformer using shifted windows

Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 2

2021

Show all 14 references

[9]

Mimo is all you need : A strong multi-in-multi-out base- line for video prediction

Shuliang Ning, Mengcheng Lan, Yanran Li, Chaofeng Chen, Qian Chen, Xunlai Chen, Xiaoguang Han, and Shuguang Cui. Mimo is all you need : A strong multi-in-multi-out base- line for video prediction. arXiv preprint arXiv: 2212.04655,

arXiv
[10]

Train short, test long: Attention with linear biases enables input length extrapolation

Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 3

2021 arXiv
[11]

Internimage: Exploring large-scale vi- sion foundation models with deformable convolutions.arXiv preprint arXiv:2211.05778, 2022

Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vi- sion foundation models with deformable convolutions.arXiv preprint arXiv:2211.05778, 2022. 2

2022 arXiv
[12]

Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion

Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3101–3109, 2021. 2

2021
[13]

Visual point cloud forecasting enables scalable autonomous driving

Zetong Yang, Li Chen, Yanan Sun, and Hongyang Li. Visual point cloud forecasting enables scalable autonomous driving. arXiv preprint arXiv:2312.17655, 2023. 1, 2, 3, 4

2023 arXiv
[14]

Radocc: Learning cross-modality occupancy knowledge through ren- dering assisted distillation

Haiming Zhang, Xu Yan, Dongfeng Bai, Jiantao Gao, Pan Wang, Bingbing Liu, Shuguang Cui, and Zhen Li. Radocc: Learning cross-modality occupancy knowledge through ren- dering assisted distillation. In Proceedings of the AAAI Con- ference on Artificial Intelligence, pages 7060–70...

2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

nuplan: A closed-loop ml-based plan- ning benchmark for autonomous vehicles

Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based plan- ning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 3

arXiv 2021

[2] [2]

Openscene: The largest up-to- date 3d occupancy prediction benchmark in autonomous driving

OpenScene Contributors. Openscene: The largest up-to- date 3d occupancy prediction benchmark in autonomous driving. https://github.com/OpenDriveLab/ OpenScene, 2023. 1, 3

work page 2023

[3] [3]

Faceformer: Speech-driven 3d facial anima- tion with transformers

Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial anima- tion with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 18770–18780, 2022. 3

work page 2022

[4] [4]

Bevdet: High-performance multi-camera 3d object de- tection in bird-eye-view

Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object de- tection in bird-eye-view. arXiv preprint arXiv:2112.11790,

arXiv

[5] [5]

Point Cloud Forecasting as a Proxy for 4D Occu- pancy Forecasting

Tarasha Khurana, Peiyun Hu, David Held, and Deva Ra- manan. Point Cloud Forecasting as a Proxy for 4D Occu- pancy Forecasting. In CVPR, 2023. 3

work page 2023

[6] [6]

Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo

Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo. arXiv preprint arXiv:2209.10248, 2022. 2

arXiv 2022

[7] [7]

Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers

Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chong- hao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers. In European con- ference on computer vision, pages 1–18. Springer, 2022. 1

2022

[8] [8]

Swin transformer: Hierarchical vision transformer using shifted windows

Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 2

2021