CycleBEV: Regularizing View Transformation Networks via View Cycle Consistency for Bird's-Eye-View Semantic Segmentation

Jeongbin Hong; Dooseop Choi; Taeg-Hyun An; Kyounghwan An; Kyoung-Wook Min

cs.CVFeb 27, 2026

CycleBEV: Regularizing View Transformation Networks via View Cycle Consistency for Bird's-Eye-View Semantic Segmentation

Authors: Jeongbin Hong, Dooseop Choi, Taeg-Hyun An, Kyounghwan An, Kyoung-Wook Min

Organizations: Electronics and Telecommunications Research Institute (ETRI) · University of Science and Technology (UST)

Abstract

Transforming image features from perspective view (PV) space to bird's-eye-view (BEV) space remains challenging in autonomous driving due to depth ambiguity and occlusion. Although several view transformation (VT) paradigms have been proposed, the challenge still remains. In this paper, we propose a new regularization framework, dubbed CycleBEV, that enhances existing VT models for BEV semantic segmentation. Inspired by cycle consistency, widely used in image distribution modeling, we devise an inverse view transformation (IVT) network that maps BEV segmentation maps back to PV segmentation maps and use it to regularize VT networks during training through cycle consistency losses, enabling them to capture richer semantic and geometric information from input PV images. To further exploit the capacity of the IVT network, we introduce two novel ideas that extend cycle consistency into geometric and representation spaces. We evaluate CycleBEV on four representative VT models covering three major paradigms using the large-scale nuScenes dataset. Experimental results show consistent improvements -- with gains of up to 0.74, 4.86, and 3.74 mIoU for drivable area, vehicle, and pedestrian classes, respectively -- without increasing inference complexity, since the IVT network is used only during training. The implementation code is available at https://github.com/JeongbinHong/CycleBEV.

Figures & tables

Figure 1: Visualization of model architectures. (a) CVTM [ 34 ] , (b) FocusBEV [ 36 ] , and (c) Proposed. Modules outside the green boxes are used only during training. Note that we exclude image feature extraction modules for simplicity. While CVTM and FocusBEV integrate the BEV2PV module, increasing computational cost and network size, the proposed approach employs it only during training. Furthermore, CVTM enforces CC in feature space (semantically vague), and FocusBEV applies feature-space projections without an explicit CC loss (semantically vague and unenforced). As a result, BEV predictions are not directly constrained. In contrast, we enforce semantic-level BEV→PV consistency using a training-only IVT that directly regularizes BEV predictions.

Figure 2: Illustration of the proposed dual-branch IVT network architecture.

Figure 3: Visualization of the proposed regularization framework.

Model	Driv.	Veh.	Ped.	Avg.
CVT	76.80	31.41	10.89	39.70
CVT + [ 34 ]	77.06 0.26↑	31.61 0.2↑	11.19 0.3↑	39.95 0.25↑
CVT + [ 36 ]	76.52 0.28↓	31.23 0.18↓	10.73 0.16↓	39.49 0.21↓
CVT+Ours	77.40 0.6↑	34.24 2.83↑	13.69 2.8↑	41.78 2.08↑
PETRv2	78.80	31.51	8.31	39.54
PETRv2 + [ 34 ]	79.45 0.65↑	31.49 0.02↓	8.91 0.6↑	39.95 0.41↑

Table 1: Quantitative comparison. The bold and underline indicate the best and second-best performance, respectively. The values in the subscript indicate the performance gain over the baseline.

\dagger

indicates that the training result could not be made comparable to others despite careful tuning.

Figure 4: Prediction results. (a) Input images and their corresponding ground-truth BEV maps, (b) BEV map prediction results. In (b), the first row shows the predictions from the four baseline models. The second, third, and fourth rows show the results when CVTM [ 34 ] , FocusBEV [ 36 ] , and Ours are applied to the baseline models, respectively. Drivable area , vehicle , and pedestrian are color-coded with gray, blue, and red, respectively. Please zoom in for better visibility.

Model	VCC	Height	Align	Driv.	Veh.	Ped.	Avg.
CVT				76.80	31.41	10.89	39.70
	✔			76.99	32.65	12.00	40.55
	✔	✔		77.12	33.76	13.32	41.40
	✔	✔	✔	77.40	34.24	13.69	41.78
	✔	✔	✔	77.23	34.14	13.65	41.67
BEVFormer				78.06	33.23	11.70	41.00

Table 2: Ablation study on the effectiveness of our contributions. VCC , Height , and Align refer to view cycle consistency, height-aware geometric regularization, and cross-view latent consistency, respectively. Results highlighted in gray are produced using a single-branch IVT network.

Model	Vis. $\geq$ 40 $\%$		Vis. $<$ 40 $\%$
Model	Veh.	Ped.	Veh.	Ped.
CVT	33.15	11.40	8.53	2.42
CVT + [ 34 ]	33.43 0.28↑	11.74 0.34↑	8.49 0.04↓	2.53 0.11↑
CVT+Ours	36.69 3.54↑	14.32 2.92↑	9.07 0.54↑	2.78 0.36↑
BEVFormer	35.32	12.40	9.01	2.60
BEVFormer + [ 34 ]	35.31 0.01↓	12.45 0.05↑	8.90 0.11↓	2.60 0.0

Table 3: Prediction performance on high and low visible objects. The values in the subscript indicate the performance gain over the baseline.

\Delta_{\mathbf{C-B}}

and

\Delta_{\mathbf{CO-B}}

respectively denote the gains of CVT and CVT+Ours over BEVFormer.

Figure 5: Prediction examples on a scene with occluded vehicles. (a) Input images (the first column), ground-truth PV segmentation maps (the second column), and PV segmentation maps predicted by the proposed IVT network (the third column). (b) Ground-truth BEV map (the first column), BEV map predicted by BEVFormer (the second column), and BEV map predicted by BEVFormer+Ours (the third column). The green boxes indicate the AV.

Model	Driv.	Veh.	Ped.	Avg.
CVT	76.80	31.41	10.89	39.70
CVT+AE	77.10 0.3↑	31.61 0.2↑	10.93 0.04↑	39.88 0.18↑
CVT+IVT	77.40 0.6↑	32.05 0.64↑	11.97 1.08↑	40.47 0.77↑
BEVFormer	78.06	33.23	11.70	41.00
BEVFormer+AE	78.06	33.30 0.07↑	11.75 0.05↑	41.04 0.04↑
BEVFormer+IVT	78.18 0.12↑	33.25 0.02↑	11.81 0.11↑	41.08 0.08↑

Table 4: Comparison with BEV auto-encoder (AE) supervision. The values in the subscript indicate the performance gain over the baseline.

Model	Driv.	Veh.	Ped.	Avg.
CVT	76.80	31.41	10.89	39.70
CVT+SA-I	76.40 0.4↓	31.75 0.34↑	11.43 0.54↑	39.86 0.16↑
CVT+SA-II	75.08 1.72↓	31.40 0.01↓	10.92 0.03↑	38.52 1.18↓
CVT+Ours	77.40 0.6↑	34.24 2.83↑	13.69 2.8↑	41.78 2.08↑
BEVFormer	78.06	33.23	11.70	41.00
BEVFormer+SA-I	78.68 0.62↑	33.70 0.47↑	12.57 0.87↑	41.65 0.65↑

Table 5: Comparison with semantic-aware (SA) image backbone training. ‘SA-I’: backbones optimized with PV segmentation; ‘SA-II’: backbones optimized with PV segmentation and irrelevant features discarded via the PV segmentation. The values in the subscript indicate the performance gain over the baseline.

Model	Aug.	CycleBEV	Driv.	Veh.	Ped.	Avg.
CVT			76.80	31.41	10.89	39.70
	✔		77.27 0.47↑	31.73 0.32↑	11.08 0.19↑	40.03 0.33↑
		✔	77.40 0.6↑	34.24 2.83↑	13.69 2.8↑	41.78 2.08↑
	✔	✔	77.43 0.63↑	34.38 2.97↑	13.59 2.7↑	41.80 2.10↑
BEVFormer			78.06	33.23	11.70	41.00
	✔		79.72 1.66↑	35.12 1.89↑	13.40 1.70↑	42.75 1.75↑

Table 6: Compatibility with typical image augmentation. The values in the subscript indicate the performance gain over the baseline.

Figure 6: Visualization of multi-view images (first column), the GT PV segmentation maps (second column), and the pseudo labels generated by Mask2Former (third column). Drivable area , vehicle , and pedestrian are color-coded with gray, blue, and red, respectively.

Model	Driv.	Veh.	Ped.	Avg.
Single-branch	82.32	70.48	32.21	61.67
Dual-branch	81.82	68.91	28.27	59.67

Table 7: PV segmentation performance of the proposed IVT network on nuScenes validation set.

Model	VCC	Height	Align	Driv.	Veh.	Ped.	Avg.
CVT				76.80	31.41	10.89	39.70
	✔			77.20	32.84	12.73	40.90
	✔			76.99	32.65	12.00	40.55
	✔	✔		77.11	33.62	13.16	41.30
	✔	✔		77.12	33.76	13.32	41.40
	✔	✔	✔	77.23	34.14	13.65	41.67

Table 8: Ablation study on the effectiveness of our contributions. VCC , Height , and Align refer to view cycle consistency, height prediction, and intermediate BEV feature alignment, respectively. Results highlighted in blue and red are produced using a single - and dual -branch IVT networks, respectively.

Figure 7: Examples of PV segmentation map prediction results. (a) GT BEV maps (left) and PV map predictions from the dual-branch IVT network (right). (b) GT PV maps (left) and their corresponding multi-view images (right). Drivable area , vehicle , and pedestrian are color-coded with gray, blue, and red, respectively.

Model	$N_{p}$	Driv.	Veh.	Ped.	Avg.
BEVFormer-S	0	78.06	33.23	11.70	41.00
BEVFormer-T	1	76.55	33.50	12.36	40.80
	2	76.55	33.47	12.69	40.90
	3	75.65	33.72	12.41	40.59
$\Delta_{\text{Ours}}$	1	0.1	0.95	1.34	0.8
	2	-0.3	0.93	1.36	0.67

Table 9: Performance gain on BEVFormer with the temporal setting.

N_{p}

denotes the number of previous frames the models take as input.

\Delta_{\text{Ours}}

denotes the performance gain achieved by BEVFormer-T+Ours over BEVFormer-T.

Model	$N_{p}$	Driv.	Veh.	Ped.
BEVFormer-S	0	78.06	33.23	11.70
$\Delta_{\text{Temp}}$	1	-1.51	0.27	0.66
	2	-1.51	0.24	0.99
	3	-2.41	0.49	0.71
$\Delta_{\text{Ours}}$	0	0.14	1.23	1.69

Table 10: Comparison with Temporal Information Aggregation.

N_{p}

denotes the number of previous frames the models take as input. The values in blue are the gains achieved by BEVFormer-T over BEVFormer-S. The values in red are the gains achieved by BEVFormer-S+Ours over BEVFormer-S.

Figure 8: Prediction examples on a scene with occluded pedestrians. (a) Input images (the first column), ground-truth PV segmentation maps (the second column), and PV segmentation maps predicted by the proposed IVT network (the third column). (b) Ground-truth BEV map (the first column), BEV map predicted by BEVFormer (the second column), and BEV map predicted by BEVFormer+Ours (the third column). The green boxes indicate the AV. Please zoom in for better visibility.

Figure 9: Prediction results. (a) Input images and their corresponding ground-truth BEV maps, (b) BEV map prediction results. In (b), the first row shows the predictions from the four baseline models. The second, third, and fourth rows show the results when CVTM [ 34 ] , FocusBEV [ 36 ] , and Ours are applied to the baseline models, respectively. Drivable area , vehicle , and pedestrian are color-coded with gray, blue, and red, respectively. Please zoom in for better visibility.

Figure 10: Prediction results. (a) Input images and their corresponding ground-truth BEV maps, (b) BEV map prediction results. In (b), the first row shows the predictions from the four baseline models. The second, third, and fourth rows show the results when CVTM [ 34 ] , FocusBEV [ 36 ] , and Ours are applied to the baseline models, respectively. Drivable area , vehicle , and pedestrian are color-coded with gray, blue, and red, respectively. Please zoom in for better visibility.

Bird's-Eye-View Segmentation

See all 11 papers

BEV-Denoise: Learning Intrinsic Noise for Accurate Bird's-Eye-View Semantic Segmentation
Jun 22, 2026Dooseop Choi, Kyounghwan An, Kyoung-Wook MinBird's-Eye-View Representation Bird's-Eye-View Segmentation
Electronics and Telecommunications Research Institute (ETRI) · University of Science and Technology (UST)
Variational Inference for Bird's Eye View Segmentation in Autonomous Driving
Jul 16, 2026Jingyue Shi, Huaicheng Li, Junhui Zhao +1Multi-View Learning Bird's-Eye-View Representation
School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100044, China · School of Information Science and Engineering, Southeast University, Nanjing 210096, China
Faster Training, Fewer Labels: Self-Supervised Pretraining for Fine-Grained BEV Segmentation
Feb 20, 2026Daniel Busch, Christian Bohn, Thomas Kurbiel +3Bird's-Eye-View Segmentation Autonomous Driving Perception
University of Wuppertal · APTIV

Jun 22, 2026J/K move · Enter open · S save

BEV-Denoise: Learning Intrinsic Noise for Accurate Bird's-Eye-View Semantic Segmentation

Dooseop Choi, Kyounghwan An, Kyoung-Wook Min

Electronics and Telecommunications Research Institute (ETRI) · University of Science and Technology (UST)

Bird's-Eye-View Representation Bird's-Eye-View Segmentation Semantic Segmentation

Open paper AskPDF arXiv

In this paper, we present a framework dubbed BEV-Denoise that estimates and removes intrinsic noise from learned Bird's-Eye-View (BEV) features to achieve accurate BEV semantic segmentation. Inspired by the noise estimation capability of Denoising Diffusion Probabilistic Models (DDPM), we design a UNet-based noise estimation module that learns to estimate the noise from the learned BEV features. The estimated noise is then subtracted from the BEV features and fed to BEV map decoders for the final prediction results. To facilitate supervision for the noise estimation module, we follow a sequential learning paradigm called Task Decomposition (TD) where a pre-trained BEV map autoencoder is employed to train a view transformation (VT) encoder. We share three key insights learned from our intensive experiments that are critical for improved performance. We apply our framework to four existing models, encompassing the three major VT paradigms. Experimental results on a large-scale real-world dataset, nuScenes, demonstrate the effectiveness of our framework.

Abstract

Bird's-Eye-View Segmentation

Explore similar work

BEV-Denoise: Learning Intrinsic Noise for Accurate Bird's-Eye-View Semantic Segmentation

Variational Inference for Bird's Eye View Segmentation in Autonomous Driving

Faster Training, Fewer Labels: Self-Supervised Pretraining for Fine-Grained BEV Segmentation