BEV Injection Connector

• The BEV injection connector is a functional module that converts perspective-view features into BEV representations via coordinate transfer, alignment, selective writing, and downstream dispatch.
• It comprises several architectural families including shared latent interfaces, geometric lift-project-pool, query-driven cross-attention, and pose-guided alignment mechanisms.
• Its design emphasizes efficiency and robustness through sparse gating, dense sampling, and dynamic supervision, enabling effective multi-modal fusion under various operational constraints.

A “BEV injection connector” (Editor’s term) denotes the interface that transfers information into, within, or out of a Bird’s-Eye View representation. Recent work does not define the term as a formal module name, but the underlying role is explicit across the literature: the connector may be a view transformer between perspective-view features and BEV queries, a lift-splat or sampling operator that writes image evidence into a ground-plane grid, a shared BEV bottleneck that fans out to task heads, a pose-guided alignment mechanism between camera and lidar BEVs, a graph-weighted cross-attention block over heterogeneous cameras, or a projector that converts BEV tensors into language-model tokens (Ma et al., 2022, Li et al., 2024, Zhao et al., 2024, Hosseinzadeh et al., 2024, Zhang et al., 28 Oct 2025, Monninger et al., 6 Mar 2026).

1. Conceptual scope and functional definition

The most stable interpretation of the connector is functional rather than taxonomic. In the survey literature, the closest terms are “view transformation,” “PV-to-BEV transformation,” “lifting 2D features to 3D space,” “view transformer,” “cross attention between BEV queries and image features,” “BEV pooling,” and “fusion under BEV” (Ma et al., 2022). In that sense, the connector is the locus at which perspective-view evidence becomes BEV-structured evidence, or at which an existing BEV tensor is dispatched to another spatial, temporal, task, or semantic domain.

This role appears in several distinct places. In multitask perception, QuadBEV uses “four distinct task-specific heads, each connected to a shared BEV feature map,” with the heads arranged “in a round-robin configuration without specialized routing architectures for different tasks”; the practical connector is therefore the shared BEV bottleneck plus the per-head convolutional encoders (Li et al., 2024). In language-grounded driving, BEVLM uses a lightweight MLP projector to map a downsampled BEV grid into a sequence of LLM-compatible tokens, making the connector a BEV-to-language interface rather than a PV-to-BEV transformer (Monninger et al., 6 Mar 2026). In collaborative vehicular perception, BEVCooper places the connector at the intermediate BEV feature level, where selected neighboring vehicles transmit compressed BEV features that are injected into the ego vehicle’s BEV fusion pipeline under deadline and bandwidth constraints (Hou et al., 22 Dec 2025).

This suggests that a BEV injection connector is best understood as a role defined by four recurring operations: coordinate transfer, alignment, selective write or read, and downstream dispatch. The specific module class varies widely.

2. Principal architectural families

Across the recent literature, connector designs fall into several recurring families.

Family	Connector operation	Representative systems
Shared latent interface	One BEV tensor branches to multiple downstream heads	QuadBEV (Li et al., 2024)
Geometric lift-project-pool	Image features are lifted with depth or sampling geometry and written into BEV	SA-BEV (Zhang et al., 2023), GeoBEV (Zhang et al., 2024)
Query-driven cross-attention	BEV queries or pillars sample image features at projected locations	FocusBEV (Zhao et al., 2024), MIC-BEV (Zhang et al., 28 Oct 2025)
Calibration-free or weak-geometry mapping	Learned decoupling or similarity fusion replaces explicit calibration-heavy lifting	CBR (Fan et al., 2023)
Pose-guided cross-modal alignment	Camera-BEV and lidar-BEV are aligned in a shared frame before fusion	BEVPose (Hosseinzadeh et al., 2024)
External BEV projection	BEV features are projected to language tokens or transmitted across agents	BEVLM (Monninger et al., 6 Mar 2026), BEVCooper (Hou et al., 22 Dec 2025)

Geometric lift-project-pool connectors remain the most direct descendants of the view-transformer abstraction described in the survey. SA-BEV keeps the LSS/BEVDepth-style lifting pipeline but inserts semantic-aware hard gating before projection, while GeoBEV replaces sparse pseudo-point pooling with Radial-Cartesian BEV Sampling, a dense radial-to-Cartesian transformation (Zhang et al., 2023, Zhang et al., 2024). Query-driven connectors instead let BEV locations pull image evidence. FocusBEV uses column-wise cross-attention with a BEV-PV-BEV cycle, and MIC-BEV uses BEV pillars, deformable cross-attention, and relation-aware graph fusion over cameras (Zhao et al., 2024, Zhang et al., 28 Oct 2025).

A different lineage reduces reliance on explicit camera geometry. CBR learns a calibration-free BEV representation for infrastructure perception by decoupling perspective-view features into front-view and BEV latents and then enhancing BEV through similarity-based cross-view fusion restricted to the same $x$-column (Fan et al., 2023). BEVPose occupies an intermediate position: it still constructs camera BEV and lidar BEV with calibrated geometry, but the key connector is the pose-guided alignment objective that enforces latent consistency in a shared BEV frame before cross-attention fusion (Hosseinzadeh et al., 2024).

3. Core mechanisms of injection

The connector’s internal mechanics differ sharply across families, but the dominant strategies are sparse gating, dense radial sampling, cycle-calibrated attention, and graph-weighted multi-view aggregation.

In SA-BEV, the connector is an explicit selective write interface. Standard lifted point features are defined as $\boldsymbol{p}_d=\alpha_d\boldsymbol{c}$, where $\alpha_d$ is a predicted depth score and $\boldsymbol{c}$ is the image context feature. SA-BEV then filters the lifted points by both depth and semantic foreground score: $$\hat{\boldsymbol{p}}_d=\mathcal{F}(\alpha_d, T_D)\mathcal{F}(\beta, T_S)\boldsymbol{p}_d.$$ Only surviving virtual points are projected into BEV pillars. With $T_D=0.0085$ and $T_S=0.25$, only $1.80\%$ of virtual points remain, yet detection improves, which makes the connector a sparse, semantic-conditioned, depth-aware BEV write mechanism rather than a dense projector (Zhang et al., 2023).

GeoBEV formalizes a denser connector. It first forms a radial BEV tensor by multiplying transposed image features and depth scores,

$$\hat{\mathbf{B}}_R=\hat{\mathbf{I}}\hat{\mathbf{D}},$$

with $$\hat{\mathbf{B}}_R \in \mathbb{R}^{W \times C \times D}$$, then converts it into a Cartesian BEV by bilinear sampling. The paper positions this as an efficient high-resolution feature transformation that avoids the vacant-cell behavior of pseudo-point pooling. In the reported ablation, RC-Sampling alone raises performance from $\boldsymbol{p}_d=\alpha_d\boldsymbol{c}$0 mAP / $\boldsymbol{p}_d=\alpha_d\boldsymbol{c}$1 NDS to $\boldsymbol{p}_d=\alpha_d\boldsymbol{c}$2 mAP / $\boldsymbol{p}_d=\alpha_d\boldsymbol{c}$3 NDS, and the full system with geometric supervision reaches $\boldsymbol{p}_d=\alpha_d\boldsymbol{c}$4 mAP / $\boldsymbol{p}_d=\alpha_d\boldsymbol{c}$5 NDS (Zhang et al., 2024).

FocusBEV implements the connector as a self-calibrated cycle view transformation. For each FPN level, a decoder-only transformer maps perspective-view features into polar BEV features, maps the rough BEV back into image space to create BEV-focused PV features, and then reinjects those calibrated PV features into BEV with residual addition: $\boldsymbol{p}_d=\alpha_d\boldsymbol{c}$6 The module then resamples polar BEV into Cartesian BEV and fuses temporally aligned history in BEV space. This connector is therefore not only a projector but also a relevance filter and a temporal BEV-to-BEV injector (Zhao et al., 2024).

MIC-BEV extends the query-driven design to heterogeneous roadside cameras. For each BEV cell $\boldsymbol{p}_d=\alpha_d\boldsymbol{c}$7, it builds a vertical pillar of $\boldsymbol{p}_d=\alpha_d\boldsymbol{c}$8 3D anchors, projects them into each camera, samples per-camera image evidence with deformable attention,

$\boldsymbol{p}_d=\alpha_d\boldsymbol{c}$9

and then fuses visible views with graph-predicted weights,

$\alpha_d$0

The weights $\alpha_d$1 are computed from a bipartite camera-to-BEV graph whose edge attributes encode normalized position offsets, range, yaw relation, and pitch. In the ablation, replacing regular spatial cross-attention with ReSCA increases M2I-Normal performance from $\alpha_d$2 to $\alpha_d$3 mAP and from $\alpha_d$4 to $\alpha_d$5 NDS (Zhang et al., 28 Oct 2025).

A complementary case is BroadBEV, whose connector is two-stage: Point-scattering injects LiDAR BEV distribution into the camera branch during BEV formation, and ColFusion exchanges self-attention structures across camera and LiDAR BEV features. The full system improves map construction from $\alpha_d$6 to $\alpha_d$7 mIoU in the reported ablation, showing that BEV injection may occur both before and after BEV formation (Kim et al., 2023).

4. Supervision, optimization, and the shaping of connector behavior

Connector quality is not determined only by forward architecture. Several systems treat the connector as a supervision-sensitive bottleneck and devote most of their novelty to how that bottleneck is trained.

QuadBEV is explicit on this point. The system pretrains the shared extractor with map segmentation only for $\alpha_d$8 epochs at learning rate $\alpha_d$9 and weight decay $\boldsymbol{c}$0, then freezes backbone, depth estimator, and BEV encoder while warming up all heads in a round-robin schedule for $\boldsymbol{c}$1 epochs per task with primary learning rate $\boldsymbol{c}$2 and auxiliary learning rate $\boldsymbol{c}$3, and finally unfreezes the full model for $\boldsymbol{c}$4 end-to-end epochs with main learning rate $\boldsymbol{c}$5, backbone learning rate $\boldsymbol{c}$6, and GradNorm-based dynamic task weighting (Li et al., 2024). The pretraining-task ablation reports the best “discount” for map pretraining, $\boldsymbol{c}$7, versus $\boldsymbol{c}$8 for detection, $\boldsymbol{c}$9 for lane, and $$\hat{\boldsymbol{p}}_d=\mathcal{F}(\alpha_d, T_D)\mathcal{F}(\beta, T_S)\boldsymbol{p}_d.$$0 for occupancy. A plausible implication is that dense BEV semantics often produce a more generally useful connector state than sparse supervisory signals.

GeoBEV injects geometry into the connector by redesigning depth supervision. Instead of LiDAR surface-only depth labels, it introduces the In-Box Label,

$$\hat{\boldsymbol{p}}_d=\mathcal{F}(\alpha_d, T_D)\mathcal{F}(\beta, T_S)\boldsymbol{p}_d.$$1

and replaces softmax-normalized one-hot depth classification with sigmoid-based occupancy-style supervision and the Centroid-Aware Inner Loss. The latter reweights positive samples by a centroid-aware factor $$\hat{\boldsymbol{p}}_d=\mathcal{F}(\alpha_d, T_D)\mathcal{F}(\beta, T_S)\boldsymbol{p}_d.$$2, emphasizing pseudo-points nearer the interior of the 3D box. In the ablation, the In-Box design with CAI yields a further gain beyond vanilla In-Box, and the combined RC-Sampling + In-Box system improves by $$\hat{\boldsymbol{p}}_d=\mathcal{F}(\alpha_d, T_D)\mathcal{F}(\beta, T_S)\boldsymbol{p}_d.$$3 mAP and $$\hat{\boldsymbol{p}}_d=\mathcal{F}(\alpha_d, T_D)\mathcal{F}(\beta, T_S)\boldsymbol{p}_d.$$4 NDS over the BEVDepth baseline (Zhang et al., 2024).

BEVPose shifts supervision from dense map labels to pose. Its alignment loss is

$$\hat{\boldsymbol{p}}_d=\mathcal{F}(\alpha_d, T_D)\mathcal{F}(\beta, T_S)\boldsymbol{p}_d.$$5

The connector is trained to maximize camera–lidar BEV similarity under correct pose and suppress it under negative poses. The resulting pretraining is data-efficient: the method reports $$\hat{\boldsymbol{p}}_d=\mathcal{F}(\alpha_d, T_D)\mathcal{F}(\beta, T_S)\boldsymbol{p}_d.$$6 mIoU with full fine-tuning and remains comparable to BEVFusion with less than $$\hat{\boldsymbol{p}}_d=\mathcal{F}(\alpha_d, T_D)\mathcal{F}(\beta, T_S)\boldsymbol{p}_d.$$7 of the annotated data; with $$\hat{\boldsymbol{p}}_d=\mathcal{F}(\alpha_d, T_D)\mathcal{F}(\beta, T_S)\boldsymbol{p}_d.$$8 labels it reaches $$\hat{\boldsymbol{p}}_d=\mathcal{F}(\alpha_d, T_D)\mathcal{F}(\beta, T_S)\boldsymbol{p}_d.$$9 mIoU (Hosseinzadeh et al., 2024).

FocusBEV similarly treats connector learning as a loss-design problem. In addition to weighted BCE and an uncertainty term, it introduces an occupancy-agnostic IoU loss over all semantic classes and reports $T_D=0.0085$0 mIoU on nuScenes and $T_D=0.0085$1 mIoU on Argoverse (Zhao et al., 2024). In calibration-free infrastructure perception, CBR uses boxes-induced foreground supervision for both the front-view and BEV branches, explicitly to “guide the view transformation and encourage the module focus on foreground objects” (Fan et al., 2023).

5. Operational constraints: calibration, efficiency, heterogeneity, and communication

Connector design is strongly conditioned by deployment constraints. Three constraints recur: camera model fidelity, compute budget, and transport budget.

The mixed pinhole–fisheye benchmark makes the first constraint explicit. There, the connector is identified with the view transformation module. Forward-projection BEVDet, backward-projection BEVFormer, and projection-free PETR react differently to fisheye distortion. In zero-shot transfer from nuScenes-trained pinhole models to mixed-optics KITTI-360, PETR and BEVDet collapse to near-zero mAP and BEVFormer reaches only $T_D=0.0085$2 mAP, while distortion-aware MEI-based projection and polar-coordinate representations recover part of the loss (Liu et al., 29 Mar 2026). The same benchmark reports that projection-free architectures are “inherently more robust and effective against fisheye distortion than other VTMs.” This suggests that the more a connector depends on exact geometric resampling, the more it benefits from explicit camera-model adaptation.

CBR addresses a different calibration problem: infrastructure cameras with unstable postures. It removes calibration parameters and additional depth supervision, decouples perspective-view features into front-view and BEV latents with two MLPs, and fuses them by same-$T_D=0.0085$3-column similarity. On DAIR-V2X, it is reported to surpass ImVoxelNet once rotation noise exceeds $T_D=0.0085$4 for BEV detection and approximately $T_D=0.0085$5 for 3D detection (Fan et al., 2023). MIC-BEV handles heterogeneity rather than calibration removal: it supports a variable number of cameras with heterogeneous intrinsic and extrinsic parameters, padding to a maximum camera count with dummy zero images and identity calibration matrices that are later excluded by invalid visibility (Zhang et al., 28 Oct 2025).

Efficiency pressures often force simplification at the connector. QuadBEV computes BEV once and reuses it for four tasks, reducing cost from the “Baseline” $T_D=0.0085$6 GFLOPs and $T_D=0.0085$7 ms to “Quad” $T_D=0.0085$8 GFLOPs and $T_D=0.0085$9 ms while keeping task scores competitive (Li et al., 2024). SA-BEV reduces active virtual points to $T_S=0.25$0 under its best threshold setting, which implicitly reduces the write burden of the connector (Zhang et al., 2023). GeoBEV reports $T_S=0.25$1 FPS for RC-Sampling at $T_S=0.25$2 BEV with depth-score downsample factor $T_S=0.25$3, versus $T_S=0.25$4 FPS for BEVPoolv2 and $T_S=0.25$5 FPS for DFA3D in the matched ablation (Zhang et al., 2024). MIC-BEV reports $T_S=0.25$6 FPS on NVIDIA L40S, close to BEVFormer’s $T_S=0.25$7 FPS despite the additional graph-enhanced fusion stage (Zhang et al., 28 Oct 2025).

Communication constraints create yet another connector variant. BEVCooper operates at the intermediate BEV feature level, where each selected collaborator transmits a locally extracted BEV feature of average size $T_S=0.25$8 KB rather than a $T_S=0.25$9 MB camera image. Its utility-driven collaborator selection and deadline-aware compression are designed around the connector rather than inside the backbone, and the reported system improves BEV perception accuracy by up to $1.80\%$0, reduces end-to-end latency by $1.80\%$1, and adds only $1.80\%$2 computational overhead (Hou et al., 22 Dec 2025).

6. Variants, misconceptions, and a unifying interpretation

A common misconception is that the connector must always be a dense camera-to-BEV rasterizer. The literature is broader. YOLO-BEV does not use explicit geometric projection at all; it arranges eight surround-view images in a $1.80\%$3 matrix with the center left blank, rotates the last row by $1.80\%$4, and uses a custom YOLO head, CustomDetect, to map multi-scale image features directly to BEV object coordinates, orientation angles, and confidence scores (Liu et al., 2023). The “Deep Perspective Transformation” model is even narrower: it consumes a single front-facing RGB image and one selected perspective-view bounding box, then regresses that vehicle’s BEV bounding box. Its output is a sparse, per-object localization in a $1.80\%$5 BEV image at $1.80\%$6 pixels per meter over a $1.80\%$7 m range, not a dense BEV tensor (Mahyar et al., 2023).

Another misconception is that the connector always terminates at a perception head. BEVLM turns the BEV grid itself into a semantic interface for LLMs by max-pooling the native BEV to $1.80\%$8, flattening it into $1.80\%$9 tokens, projecting each token with a sequential $$\hat{\mathbf{B}}_R=\hat{\mathbf{I}}\hat{\mathbf{D}},$$0LayerNorm, Linear, GELU, Linear$$\hat{\mathbf{B}}_R=\hat{\mathbf{I}}\hat{\mathbf{D}},$$1 MLP, and bracketing the token span with <bev> and </bev> (Monninger et al., 6 Mar 2026). The connector here is a dense BEV-grid-to-token-sequence projector. Its role is not detection or segmentation but cross-view reasoning and semantic distillation from a frozen LLM back into the BEV encoder.

The accumulated evidence therefore supports a unifying interpretation. A BEV injection connector is not a single operator class; it is the structurally privileged interface at which BEV becomes the medium of exchange. That exchange may link perspective cameras to a ground-plane grid, one BEV tensor to another across time or agents, a shared BEV bottleneck to multiple tasks, or a BEV scene representation to language. This suggests three recurrent design principles. First, connectors become more effective when they preserve or explicitly restore spatial frame consistency, whether through calibration, pose, or ego-motion. Second, connector quality improves when the write path is selective rather than indiscriminate, as in semantic filtering, relation-aware fusion, or collaborator selection. Third, supervision placed at the connector—depth, occupancy, foreground masks, pose consistency, or task balancing—often matters as much as the forward operator itself.