BEV-ODOM2: Enhanced BEV-based Monocular Visual Odometry with PV-BEV Fusion and Dense Flow Supervision for Ground Robots

Yufei Wei; Chenxiao Hu; Wangtao Lu; Sha Lu; Yuxiang Cui; Fuzhang Han; Rong Xiong; Yue Wang

cs.ROSep 18, 2025

BEV-ODOM2: Enhanced BEV-based Monocular Visual Odometry with PV-BEV Fusion and Dense Flow Supervision for Ground Robots

Authors: Yufei Wei, Chenxiao Hu, Wangtao Lu, Sha Lu, Yuxiang Cui, Fuzhang Han, Rong Xiong, Yue Wang

Abstract

Scale-consistent ego-motion estimation is fundamental for autonomous ground robots. Bird's-Eye-View (BEV) representation naturally addresses the scale drift problem of monocular visual odometry (MVO) by providing a metric-scaled planar workspace, enabling the simplification of 6-DoF ego-motion to a more robust 3-DoF model. However, existing BEV-based methods suffer from two key limitations: sparse supervision signals from pose-only training, and information loss during perspective-to-BEV projection. We present BEV-ODOM2, an enhanced framework in the BEV-ODOM line that addresses both limitations without supervision modalities beyond the pose ground truth. Our approach introduces (1) pose-derived dense rigid BEV flow supervision, which reparameterizes the 3-DoF pose ground truth into a pixel-level training signal, and (2) Perspective View (PV)-BEV fusion, which computes correlation volumes before projection to retain additional motion cues and help alleviate projection ambiguity. An enhanced rotation sampling strategy further balances diverse motion patterns during training. We evaluate on four datasets with varied spatial scales: KITTI, Oxford, NCLT, and our newly collected ZJH-VO benchmark. BEV-ODOM2 reduces the RTE of BEV-ODOM by 39% on average across the four datasets. It further enables closed-loop navigation on a physical robot with centimeter-level cross-track error. Real-time inference on an NVIDIA Jetson AGX Orin confirms edge deployment feasibility. The code and the ZJH-VO dataset are publicly released to facilitate future research.

Figures & tables

Fig. 1: Comparison of BEV-based monocular odometry methods. Unlike prior approaches needing extra annotations or limited by sparse supervision, our BEV-ODOM2 leverages PV-BEV fusion and dense rigid BEV flow to provide rich supervision from pose data alone.

Fig. 2: Overview of the BEV-ODOM2 framework. The PV-BEV Encoder shares a ResNet-50+FPN backbone and projects both PV features and PV correlation volumes into BEV space via LSS. The PV Decoder regresses a 5-DoF pose from PV correlations. The BEV Decoder fuses projected PV and BEV-native correlations through a FuseNet to jointly predict the dense rigid BEV flow and the final 3-DoF pose.

Fig. 3: Construction of the pose-derived dense rigid BEV flow ground truth. A 3-DoF relative pose (yaw and planar translation) defines a rigid transformation on the BEV grid (Eqs. 11 – 12 ), yielding a per-pixel displacement field (right) as dense supervision.

Fig. 4: Case Study of BEV-ODOM2 Performance on Four Datasets. The figure illustrates predicted trajectories for representative straight (S), left-turn (L), and right-turn (R) maneuvers. For each case, we visualize the predicted dense rigid BEV flow and its corresponding error map, with most pixel errors below 0.4. Below, violin plots show the distribution of translational and rotational errors for 51-frame sequences centered on these maneuvers.

Fig. 5: Logarithmic scale factor variation on the NCLT (top) and Oxford (bottom) datasets, demonstrating the consistently low scale drift of BEV-based methods.

Fig. 6: Sensitivity analysis on KITTI Seq10. Top : per-frame translation error in meters and inter-frame elevation change $|\Delta z|$ , with the P75 threshold marking high-elevation-change regions. Bottom : sliding-window Pearson correlation between $|\Delta z|$ and the per-frame translation error.

Fig. 7: Impact of dynamic objects on estimation accuracy. Each panel shows the median BEV flow end-point error (solid) and median per-frame translation error in meters (dashed) versus the per-frame dynamic object pixel ratio, with interquartile ranges shaded and sample counts shown as histograms.

Fig. 8: Closed-loop route following on routes T, A, S, and E. Top: taught reference (black) and the driven trajectories (LiDAR ground truth) of BEV-ODOM2 (red) and BEV-ODOM (blue) over the ten repeats, all drawn in the shared LiDAR map frame without any trajectory alignment; the green star marks each start. Bottom: RTE (%) and RRE (°/10 m) of the VO estimate against the LiDAR ground truth as a function of the sub-trajectory length, aggregated over all runs and routes.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Fig. 9: ZJH-VO Multi-Scale Dataset Overview. Left: Data collection platform and camera coordinate system setup with PV-BEV transformation. Center: Representative trajectory paths across four floors (Floor 9, 4, 1, and 0). Right: Sample image pairs from each environment showing diverse scene characteristics.

Fig. 10: ZJH-VO Multi-Scale dataset motion statistics across four environments. Left: Inter-frame motion distance and rotation angle distributions with statistical measures. Right: Mean motion velocity and total trajectory distance for each environment sequence.

Method	Tier 1: Suburban & Urban Driving						Tier 2: Campus-Level Navigation				Tier 3: Small-Scale Env.
	KITTI		Oxford				NCLT				ZJH-VO
	Seq09	Seq10	11-12	15-13	16-14	17-12	02-02	02-19	03-17	08-20	F-09	F-04	F-01	F-00
BEV-DWPVO [ 6 ]	2.03	1.45	0.39	0.41	0.36	0.50	0.22	0.20	0.20	0.21	0.08	0.07	0.09	0.07
BEV-ODOM [ 7 ]	0.83	1.04	0.58	0.46	0.61	0.53	0.20	0.16	0.17	0.21	0.05	0.04	0.04	0.06
BEV-ODOM2 (Ours)	1.01	0.89	0.36	0.33	0.33	0.37	0.17	0.13	0.14	0.16	0.04	0.04	0.03	0.05

Appendix

TABLE VII: LATERAL (CROSS-TRACK) DEAD-RECKONING ERROR @ 10 s (UNIT: m)

Method	Tier 1: Suburban & Urban Driving						Tier 2: Campus-Level Navigation				Tier 3: Small-Scale Env.
	KITTI ( $<$ 3.66 m)		Oxford ( $<$ 1.83 m)				NCLT ( $<$ 0.50 m)				ZJH-VO ( $<$ 0.25 m)
	Seq09	Seq10	11-12	15-13	16-14	17-12	02-02	02-19	03-17	08-20	F-09	F-04	F-01	F-00
BEV-DWPVO [ 6 ]	83.4	95.9	95.4	96.8	97.7	90.4	80.4	83.1	82.2	81.5	80.7	85.1	75.7	88.2
BEV-ODOM [ 7 ]	98.8	92.7	88.0	92.1	87.6	90.0	85.9	87.3	86.4	84.5	92.8	91.7	98.2	92.1
BEV-ODOM2 (Ours)	100.0	97.0	95.9	98.0	97.3	94.8	90.4	94.2	92.9	94.2	96.3	94.3	98.7	97.7

Appendix

TABLE VIII: LATERAL LOCALIZATION AVAILABILITY @ 10 s (%)

Fig. 11: Lateral localization availability degradation over GNSS outage duration (5–30 s) on four datasets. Each panel shows the mean availability (solid line) and sequence-level min–max range (shaded band) for three BEV-based methods. Background bands indicate quality tiers: Excellent ( $\geq$ 95%), Good (85–95%), Acceptable (70–85%), and Poor ( $<$ 70%). Lateral ceilings are set per scenario: 3.66 m for KITTI suburban driving, 1.83 m for Oxford urban driving, 0.50 m for NCLT campus-level navigation, and 0.25 m for ZJH-VO small-scale structured environments.

Explore similar work

MVOFormer: Flow-Semantic Transformer for Robust Monocular Visual Odometry
Jun 15, 2026Jituo Li, Shunwang Sun, Jialu Zhang +5Visual Inertial Odometry Monocular
State Key Laboratory of Fluid Power and Mechatronic Systems, Zhejiang University and with Zhejiang Key Laboratory of Industrial Big Data and Robot Intelligent Systems. They are also with the School of Mechanical Engineering, Zhejiang University, Hangzhou 310027, China, and with Robotics Institute, Zhejiang University, Hangzhou 310027, China · School of Artificial Intelligence and Robotics, Hunan University, Changsha, 410012, China · Rural Health Research Institute, Charles Sturt University, Sydney, Australia +1
Learning Ego-Centric BEV Representations from a Perspective-Privileged View: Cross-View Supervision for Online HD Map Construction
May 12, 2026Daniel Lengerer, Mathias Pechinger, Klaus Bogenberger +1Bird'S-Eye View Lane-Level Map Generation
Technical University of Applied Sciences Augsburg, Germany · Technical University of Munich, Germany
Open-Vocabulary BEV Segmentation with 3D-Aware Geometric Constraints
Jun 23, 2026Hojun Choi, Seulbin Hwang, Dae Jung Kim +3Bird'S-Eye View Open-Vocabulary 3D Segmentation
KAIST AI, South Korea · NAVER LABS, South Korea

Jun 15, 2026, cs.CVJ/K move · Enter open · S save

MVOFormer: Flow-Semantic Transformer for Robust Monocular Visual Odometry

Jituo Li, Shunwang Sun, Jialu Zhang, Xinqi Liu+4

State Key Laboratory of Fluid Power and Mechatronic Systems, Zhejiang University and with Zhejiang Key Laboratory of Industrial Big Data and Robot Intelligent Systems. They are also with the School of Mechanical Engineering, Zhejiang University, Hangzhou 310027, China, and with Robotics Institute, Zhejiang University, Hangzhou 310027, China · School of Artificial Intelligence and Robotics, Hunan University, Changsha, 410012, China · Rural Health Research Institute, Charles Sturt University, Sydney, Australia+1

Visual Inertial Odometry Monocular Sequential Visual Localization Orb-Slam2 Dual-Encoder Architectures

Open paper AskPDF arXiv

Monocular visual odometry (MVO) is foundational to autonomous navigation and robotic localization. However, existing learning-based MVO approaches often struggle with either a lack of interpretable, complementary features or overly complex multi-stage architectures. These limitations inherently restrict their robustness and cross-domain generalization. In this work, we propose MVOFormer, a novel transformer framework for robust monocular visual odometry. Our architecture features a Flow-Semantic Dual Branch Encoder that synergizes dense geometric motion cues with object-centric semantic priors, explicitly distinguishing static structures from dynamic distractors. These representations are then fused by an Iterative Multimodal Decoder, enabling coarse-to-fine pose refinement while dynamically suppressing attention on unreliable regions. Extensive evaluations demonstrate that, without any target-domain fine-tuning, MVOFormer achieves superior zero-shot generalization and robustness, significantly outperforming prior learning-based frame-to-frame methods across diverse benchmarks including TartanAir, KITTI, TUM-RGBD, and ETH3D-SLAM.

Abstract

Explore similar work

MVOFormer: Flow-Semantic Transformer for Robust Monocular Visual Odometry

Learning Ego-Centric BEV Representations from a Perspective-Privileged View: Cross-View Supervision for Online HD Map Construction

Open-Vocabulary BEV Segmentation with 3D-Aware Geometric Constraints