Autonomous driving systems increasingly rely on multimodal perception, fusing inputs from cameras, LiDAR, and other sensors to improve safety and reliability. The assumption has been that combining complementary information yields a more robust overall model. But a new paper from arXiv challenges this assumption, introducing a benchmark called UBD-Bench to study a phenomenon the authors call unimodal branch degradation.
The core finding is that a multimodal model can produce strong final predictions while one of its sensor-specific branches has become weak or degraded. In other words, the fusion process can compensate for a failing modality, masking the problem in the overall output. This is concerning because such hidden degradation may go unnoticed during evaluation, only surfacing in edge cases where the weak branch is needed most.
UBD-Bench provides a way to measure and analyze this degradation across different multimodal perception tasks. By benchmarking branch-level performance separately from fused performance, the work aims to help researchers identify when and why individual sensor branches fail. The paper's contribution is primarily diagnostic, offering a tool for more transparent evaluation of multimodal driving models rather than proposing a new fusion method itself.