Table 1 Sensor fusion models with their strengths and limitations.

Early

MEFA³¹

High-resolution fusion, fine-grained

spatial feature alignment

Rigid fusion, high computational

cost, limited real-time use

VirConv³²

Lightweight, efficient voxel feature

propagation

Limited adaptability, struggles

with semantic inconsistencies

Mid-Level

BEVFusion4D³³

Temporal BEV fusion, consistent

scene aggregation

Executes fusion regardless of

feature alignment, redundant

computations

CRN³⁴

adar-Camera feature fusion with

dynamic spatial reasoning

Lacks semantic consistency gating

, sensitive to modality disagreement

DeepStep²¹

Progressive integration, improved

temporal continuity

Unconditional fusion execution,

limited adaptability

Late

C-CLOCs³⁵

Modality-specific decision making,

robustness to misalignment

Post-hoc alignment, limited

semantic feature utilization

BAFusion¹⁷

Uncertainty-aware fusion,

reliability weighting

Unconditional fusion, limited

adaptability to sensor

disagreement

Hybrid

DecoratingFusion²⁵

Multi-stage refinement, dense

feature propagation

Executes all fusion stages

unconditionally, increased latency

MS-Occ²⁶

Multi-scale occupancy reasoning,

strong spatial coverage

Performs all fusion stages

unconditionally, computationally

intensive

HydraFusion³⁷

Flexible per-branch processing,

adaptable fusion routes

Static fusion pipelines within

each branch, unnecessary

computations

RCBEV³⁶

Radar-Camera feature alignment,

modality-specific spatial adaptation

Limited to Radar-Camera pairs,

lacks generalized modality fusion