MBFNet: Motion-guided BEV Fusion for 3D Detection

• The paper introduces MBFNet as a motion-conditioned BEV fusion module that mitigates temporal misalignment using deformable attention.
• It extracts motion features from BEV feature differences with channel-wise attention to effectively align dynamic object motions in a recurrent pipeline.
• Empirical results on nuScenes demonstrate improved detection performance and robustness under occlusion and motion blur compared to static fusion baselines.

Searching arXiv for MBFNet and closely related BEV fusion papers to ground the article in the cited literature. Motion-guided BEV Fusion Network (MBFNet) is a bird’s-eye-view temporal fusion module for multi-camera 3D perception that was introduced as the alignment core of OnlineBEV. It addresses the degradation that arises when bird’s-eye-view features are aggregated over many frames even after ego-motion compensation, because dynamic objects still move in BEV space and induce temporal misalignment. In OnlineBEV, MBFNet operates inside a recurrent temporal fusion pipeline, extracting motion features from consecutive BEV queries and using those motion cues to align historical BEV features to the current frame through deformable attention, after which the aligned history is fused with the current BEV representation for 3D detection (Koh et al., 11 Jul 2025). More broadly, MBFNet belongs to a family of BEV-centric fusion designs in which BEV is treated as the common coordinate system for combining multi-view, multi-modal, and temporal information, as seen in unified spatiotemporal fusion, moving-object segmentation, and camera–radar fusion systems (Qin et al., 2022, Zhou et al., 2023, Kim et al., 2024, Cong et al., 5 Mar 2025, Montiel-Marín et al., 12 Sep 2025).

1. Definition and problem setting

In OnlineBEV, temporal fusion is motivated by the observation that single-frame BEV pipelines fail when the current frame is degraded by occlusion or motion blur, when objects are briefly invisible or partially visible, or when long-range objects are too weak in a single frame. BEV-based temporal aggregation is attractive because BEV features are spatially aligned in a fixed ground plane and, after ego-motion compensation, features from different time steps should correspond to the same world coordinates. However, the performance gain from temporal aggregation is limited when combining a large number of image frames because dynamic changes in BEV features over time caused by object motion produce temporal misalignment (Koh et al., 11 Jul 2025).

MBFNet is designed to solve precisely that failure mode. Within OnlineBEV, it addresses temporal misalignment of BEV features due to dynamic object motion in a recurrent temporal fusion setting. The method maintains a single historical BEV feature map $H_{t-1}$ and updates it with each new frame, rather than storing many historical BEV tensors. A plausible implication is that MBFNet should be understood not merely as a fusion block but as a motion-conditioned alignment mechanism that makes recurrent BEV memory viable over long horizons (Koh et al., 11 Jul 2025).

The broader literature places this design in a recognizable context. UniFusion formulates BEV fusion as a unified spatial-temporal transformer over BEV tokens and emphasizes learnable temporal weights rather than fixed averaging (Qin et al., 2022). MotionBEV formulates moving object segmentation through explicit BEV appearance and motion branches bridged by a co-attention module (Zhou et al., 2023). CRT-Fusion similarly argues that naive temporal aggregation in BEV suffers from misalignment for moving objects and introduces motion-aware temporal fusion for radar–camera detection (Kim et al., 2024). These works collectively indicate that MBFNet is part of a shift from static BEV aggregation toward temporally adaptive, motion-aware BEV fusion.

2. Architectural role inside OnlineBEV

OnlineBEV consists of an image backbone, a Lift-Splat-Shoot view transformation to BEV, a recurrent temporal fusion block, and a CenterPoint-style BEV detection head. Multi-view images are first encoded by a CNN backbone and transformed into a BEV feature map $F_t$. The recurrent module maintains a historical BEV feature $H_{t-1}$, and at each transformer layer $l$ defines a target query $q_t^{(l)}$ initialized from the current BEV and a historical query $q_{t-1}^{(l)}$ initialized from the memory feature. MBFNet is applied at every layer to align $q_{t-1}^{(l)}$ to $q_t^{(l)}$ and then fuse the aligned history into the target query. After $L$ layers, $q_t^{(L)}$ becomes the enhanced current BEV feature $F_t$0, while $F_t$1 becomes the aligned historical feature $F_t$2 used for temporal consistency learning (Koh et al., 11 Jul 2025).

This recurrent design is contrasted with parallel temporal fusion, in which a fixed set of $F_t$3 historical BEV features is stored and aggregated at every time step. OnlineBEV instead stores only one historical BEV feature map, so memory usage is constant and the effective number of combined frames can grow through recurrence. The paper explicitly frames this as a way to increase the effective number of combined features with minimal memory usage (Koh et al., 11 Jul 2025).

The following summary captures the MBFNet placement in OnlineBEV.

Component	Function	Output
Image backbone + LSS	Multi-view image encoding and BEV generation	$F_t$4
Recurrent BEV fusion	Iterative temporal fusion over transformer layers	$F_t$5
MBFNet	Motion-guided alignment and fusion of history and current BEV	updated $F_t$6, $F_t$7
Detection head	3D object detection in BEV	3D boxes

A plausible implication is that OnlineBEV treats MBFNet as the decisive factor that allows recurrent fusion to outperform parallel fusion despite using only a single stored historical BEV tensor.

3. Internal mechanics: motion feature extraction and motion-guided warping attention

MBFNet is decomposed into two modules: the Motion Feature Extractor (MFE) and the Motion-Guided BEV Warping Attention (MGWA). At transformer layer $F_t$8, the MFE computes a motion feature from the difference between the current and historical queries:

$F_t$9

Here, the feature difference $H_{t-1}$0 captures temporal change in BEV feature space, the fully connected layer is applied per BEV token, and CWA is a channel-wise attention module in the style of SENet, used to emphasize informative motion channels and suppress noise (Koh et al., 11 Jul 2025). In the OnlineBEV ablation, replacing the baseline with a difference-only MFE improves NDS from 50.9 to 51.1, and adding channel-wise attention improves it further to 51.4, indicating that motion extraction is not reducible to a raw residual alone (Koh et al., 11 Jul 2025).

MGWA then uses the motion feature as the query to a deformable cross-attention operator over the historical BEV feature. For a BEV reference point $H_{t-1}$1, the aligned historical feature is written as

$H_{t-1}$2

This means that motion features parameterize the deformable sampling offsets and weights used to read from the historical BEV tensor. Where motion is weak, the offsets remain small and alignment is close to identity; where motion is strong, the attention can shift sampling locations to follow moving objects in BEV (Koh et al., 11 Jul 2025). This suggests an implicit flow-like mechanism: MBFNet does not regress a dense motion field explicitly, but its deformable attention implements motion-conditioned spatial warping.

After deformable attention, the aligned historical query is updated through a residual connection, layer normalization, and a feed-forward network. The target query is then updated by concatenating the current and aligned historical queries and compressing them through another feed-forward network:

$H_{t-1}$3

Because this update is applied repeatedly across transformer layers, the alignment and fusion are progressively refined rather than performed in a single step (Koh et al., 11 Jul 2025).

The mechanism has strong affinities with other motion-aware BEV designs. MotionBEV uses explicit motion features and an Appearance–Motion Co-attention Module to let motion guide BEV appearance features (Zhou et al., 2023). CRT-Fusion predicts dense BEV velocity and occupancy maps and then warps historical BEV features according to predicted motion before temporal fusion (Kim et al., 2024). MBFNet differs in that its motion signal is derived directly from feature differences and consumed by deformable attention, but the underlying principle—motion should control temporal alignment in BEV—is shared.

4. Temporal consistency supervision

OnlineBEV complements MBFNet with a Heatmap-based Temporal Consistency Loss, also called HTC-loss. The motivation is that detection losses alone do not directly penalize dense feature-level misalignment between the current BEV feature $H_{t-1}$4 and the aligned historical BEV feature $H_{t-1}$5. To enforce explicit temporal alignment, the same detection head is applied to both features, producing heatmaps $H_{t-1}$6 and $H_{t-1}$7, and the consistency term is defined as

$H_{t-1}$8

The gradient is stopped on the $H_{t-1}$9 branch, so the aligned historical feature is trained to match the detection behavior of the current feature (Koh et al., 11 Jul 2025). The total loss is

$l$0

with $l$1, $l$2, and $l$3 (Koh et al., 11 Jul 2025).

Empirically, this consistency loss provides an additional gain beyond MBFNet itself. On nuScenes validation with ResNet-50, recurrent fusion without MBFNet yields 50.4 NDS and 40.8 mAP; adding MBFNet raises the result to 51.4 NDS and 42.1 mAP; adding HTC-loss further raises it to 51.9 NDS and 42.5 mAP (Koh et al., 11 Jul 2025). This suggests that MBFNet alignment is improved when the model is explicitly required to make aligned history and current features induce similar detection heatmaps.

An analogous theme appears in other work, though with different targets. BEV-ODOM2 supervises dense BEV optical flow derived from pose to provide pixel-level motion guidance in BEV (Wei et al., 18 Sep 2025). MotionBEV combines weighted cross-entropy and Lovász-Softmax to directly optimize moving-object segmentation quality from motion-aware BEV features (Zhou et al., 2023). The general pattern is that motion-aware BEV fusion benefits from auxiliary supervision that acts on intermediate motion-aligned representations rather than only on the final task head.

5. Performance and empirical behavior

OnlineBEV reports that MBFNet substantially improves recurrent temporal fusion. On nuScenes validation with ResNet-50 and no perspective pre-training, a single-frame baseline achieves 41.9 NDS and 35.4 mAP. Recurrent fusion without MBFNet raises the result to 50.4 NDS and 40.8 mAP. Adding MBFNet raises it to 51.4 NDS and 42.1 mAP, and adding HTC-loss raises it further to 51.9 NDS and 42.5 mAP (Koh et al., 11 Jul 2025). Compared with a 17-frame parallel fusion baseline, the recurrent design with MBFNet and HTC-loss still performs better while using lower memory (Koh et al., 11 Jul 2025).

On the nuScenes test set, OnlineBEV achieves 63.9% NDS in the camera-only 3D object detection task (Koh et al., 11 Jul 2025). The paper explicitly states that this surpasses SOLOFusion while storing only one BEV feature and achieving better robustness. Under motion blur and occlusion corruptions, OnlineBEV degrades less than SOLOFusion: under motion blur, SOLOFusion drops by 10.5 mAP and 6.5 NDS, whereas OnlineBEV drops by 4.8 mAP and 3.0 NDS; under occlusion, SOLOFusion drops by 13.0 mAP and 11.1 NDS, whereas OnlineBEV drops by 8.8 mAP and 5.7 NDS (Koh et al., 11 Jul 2025).

The ablation on MFE also indicates improved velocity-sensitive behavior. When difference-only motion extraction is replaced by full MFE with channel-wise attention, mAVE improves from 0.289 to 0.287 and mATE improves from 0.613 to 0.611, alongside the NDS gain (Koh et al., 11 Jul 2025). This suggests that motion-guided temporal alignment is beneficial not only for classification heatmaps but also for state estimation on moving objects.

Related work reinforces these empirical trends. CRT-Fusion reports that motion-aware temporal fusion adds a further 1.1 NDS and 1.1 mAP after strong per-frame camera–radar fusion, and that gains are largest at medium velocities (Kim et al., 2024). MotionBEV reports that explicit BEV motion features and co-attention substantially outperform direct concatenation, with the best SemanticKITTI-MOS validation IoU reaching 76.54% for full AMCM with BEV motion features (Zhou et al., 2023). BEVMOSNet reports that camera–radar–LiDAR fusion with camera–radar deformable cross-attention followed by LiDAR concatenation reaches 62.59% mIoU for moving-vehicle BEV segmentation on nuScenes, exceeding a concatenation baseline by 2.35 mIoU (Cong et al., 5 Mar 2025). These results suggest that MBFNet’s gains are consistent with a broader pattern: explicit motion-aware BEV alignment yields measurable improvements over motion-agnostic temporal aggregation.

6. Relation to adjacent BEV fusion paradigms and extensions

MBFNet is most directly tied to recurrent camera-only temporal fusion, but its design is best understood against several neighboring paradigms. UniFusion unifies spatial and temporal fusion in a single BEV transformer and emphasizes that temporal fusion weights should be learnable rather than fixed (Qin et al., 2022). FocusBEV uses ego-motion-based temporal fusion with a memory bank for monocular BEV segmentation, but its temporal module aligns only with ego-motion and primarily benefits static layout classes (Zhao et al., 2024). A plausible implication is that MBFNet can be viewed as extending ego-motion alignment with a learned correction for dynamic object motion.

In radar–camera fusion, CRT-Fusion provides a close structural analogue. It first builds a strong per-frame fused BEV representation, then predicts per-pixel BEV velocity and occupancy, and finally warps and fuses historical BEV features recurrently according to those motion estimates (Kim et al., 2024). By contrast, CaR1 provides a BEVFusion-style camera–radar baseline with grid-wise radar encoding and adaptive BEV fusion but does not explicitly model Doppler-guided fusion or motion-specific losses (Montiel-Marín et al., 12 Sep 2025). The CaR1 authors explicitly note that their architecture is the kind of BEV fusion baseline one would extend to obtain an MBFNet-style design, by injecting radial velocity, sweep timing, and motion-conditioned fusion weights (Montiel-Marín et al., 12 Sep 2025).

For moving-object segmentation, MotionBEV and MV-MOS show two additional variants of the same core idea. MotionBEV computes explicit BEV motion features from height differences across aligned scans and uses motion-guided co-attention to refine appearance features (Zhou et al., 2023). MV-MOS fuses BEV and range-view motion features and uses a semantic branch plus a Mamba-based density-aware fusion block to preserve sparse motion cues (Cheng et al., 2024). These systems suggest that MBFNet is not tied to 3D detection; it is a more general template in which BEV motion features guide the fusion of spatial or semantic features.

A further extension appears in BEV-ODOM2, which computes a perspective-view correlation volume before projection, projects that motion-rich representation into BEV, and fuses it with a native BEV correlation volume under dense BEV flow supervision (Wei et al., 18 Sep 2025). This suggests a different but compatible route for MBFNet-style systems: motion can be estimated in a richer source domain such as perspective view and then transported into BEV for fusion.

Taken together, the literature indicates several recurrent design principles for MBFNet-style architectures: BEV as the common fusion space; motion features as first-class alignment signals; lightweight yet adaptive fusion operators such as deformable attention or channel/spatial gating; and auxiliary supervision that constrains temporal consistency or dense motion fields (Qin et al., 2022, Zhou et al., 2023, Kim et al., 2024, Cong et al., 5 Mar 2025, Koh et al., 11 Jul 2025, Wei et al., 18 Sep 2025, Montiel-Marín et al., 12 Sep 2025).

7. Limitations and research directions

OnlineBEV notes that MBFNet currently uses implicit motion features derived from BEV feature differences rather than explicit motion labels or dense velocity supervision (Koh et al., 11 Jul 2025). The paper points to explicit motion extraction as a future direction, potentially supervised by auxiliary sensors such as LiDAR or radar. This suggests that the current MBFNet formulation is powerful but only partially exploits the motion signal available in multi-sensor autonomous driving settings.

The radar-centered literature sharpens that point. BEVMOSNet emphasizes that radar points include 15 non-position attributes carrying Doppler and processed velocity information, and explicitly argues that radar velocity should guide camera features through deformable cross-attention in BEV (Cong et al., 5 Mar 2025). CRT-Fusion likewise shows that accurate motion estimation from fused radar–camera BEV is essential; applying its motion-guided temporal fusion to camera-only BEV features actually hurts performance, indicating that motion guidance is only as good as the motion features on which it rests (Kim et al., 2024). CaR1, by contrast, motivates radar as providing reliable distance and velocity measurements but does not use Doppler or radial velocity explicitly in fusion, leaving that design space open (Montiel-Marín et al., 12 Sep 2025).

Another limitation concerns the distinction between static and dynamic content. FocusBEV shows that short-window ego-motion alignment is beneficial, especially for static layouts, but dynamic object gains are smaller without explicit object-motion modeling (Zhao et al., 2024). UniFusion’s unified spatiotemporal attention can learn temporal weighting, but it does not impose an explicit motion model either (Qin et al., 2022). A plausible implication is that future MBFNet variants may separate static and dynamic BEV pathways or combine ego-motion warping with learned motion-conditioned residual alignment.

Computational scale also remains central. Full spatiotemporal attention over dense BEV tokens is expensive, which motivates deformable attention, recurrent memory, or lightweight BEV-space fusion (Qin et al., 2022, Koh et al., 11 Jul 2025). OnlineBEV reports that the recurrent fusion block with MBFNet increases GFLOPs only modestly, from 192.2 to 205.7 for the ResNet-50 setting, while reducing system memory relative to SOLOFusion from 3.9 GB to 3.4 GB (Koh et al., 11 Jul 2025). This suggests that motion-guided recurrent fusion is attractive partly because it offers a favorable accuracy–memory trade-off.

Overall, the current literature supports a concise characterization of MBFNet: a motion-conditioned BEV alignment-and-fusion module that enables recurrent temporal aggregation by compensating for dynamic object motion in BEV. Its specific instantiation in OnlineBEV is camera-only and feature-difference-driven (Koh et al., 11 Jul 2025), but adjacent work indicates natural extensions toward explicit velocity supervision, radar-guided fusion, dense motion heads, and unified spatiotemporal BEV transformers (Qin et al., 2022, Zhou et al., 2023, Kim et al., 2024, Cong et al., 5 Mar 2025, Montiel-Marín et al., 12 Sep 2025).