Research map / Detection and segmentation / 3D detection and driving

Hyperdimensional Uncertainty Quantification for Multimodal Uncertainty Fusion in Autonomous Vehicles Perception

CVPR 2025 · Detection and segmentation › 3D detection and driving · approach: LiDAR perception for driving · 9 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HyperDUM, a deterministic uncertainty method that uses channel- and patch-wise hyperdimensional projection and bundling to quantify feature-level epistemic uncertainty and reweight multimodal features before fusion in autonomous driving perception.

Problem

Existing uncertainty quantification methods for autonomous driving either quantify only task-level output uncertainty or are too computationally expensive (Bayesian approximations, ensembles), and they do not address epistemic uncertainty at the multimodal feature fusion level.

Approach

Projects latent features into hyperdimensional prototypes via channel-wise and patch-wise projection and bundling, measures similarity between sample hypervectors and context-label prototypes as epistemic uncertainty, and uses it to adaptively weight each modality's features before fusion.

Key results

• On aiMotive 3D object detection, HyperDUM reaches 66.70 all-point AP mean, a 2.01% gain over LDU (64.69).
• On DeLiVER semantic segmentation, HyperDUM achieves 67.59 mean mIoU, a 1.29 improvement over the CMNeXt baseline (66.30).
• HyperDUM uses up to 2.36x fewer FLOPs than LDU on aiMotive and up to 38.30x fewer trainable parameters.

Measured effects reported in the paper

Surprising findings

• Applying uncertainty weighting after the fusion block gives much smaller gains than pre-fusion weighting, supporting the claim that modality uncertainty propagates into fusion.
• Hyperdimension performance plateaus and then declines past about 10k dimensions, which the authors attribute to over-sparsity and overfitting.

Limitations

• Requires labels for supervised prototype bundling; semi-supervised or unsupervised variants lack theoretical guarantees.
• Under distant-region evaluation and rain conditions, uncertainty weighting can suppress features and degrade performance relative to the baseline.
• Training uses a single NVIDIA A100 GPU and experiments are limited to two datasets.

Experimental setup

Datasets: aiMotive DeLiVER

Benchmarks: aiMotive DeLiVER

Metrics: all-point AP 11-point AP mIoU ECE FLOPs trainable parameters

Baselines: BEVFusion InfMCD InfNoise PostNet LDU GeminiFusion CMNeXt

Builds on: Latent Deterministic Uncertainty BEVFusion CMNeXt Posterior Networks Monte Carlo dropout

Closest papers by meaning

• Predictive Uncertainty Quantification for Bird's Eye View Segmentation: A Benchmark and Novel Loss Function (ICLR 2025 · similarity 0.87)
• Benchmarking Uncertainty Disentanglement: Specialized Uncertainties for Specialized Tasks (NeurIPS 2024 · similarity 0.86)
• FUSE: Quantifying Uncertainty in Vision-Language Models by Bayesian Fusing Epistemic and Aleatoric Uncertainty (ICML 2026 · similarity 0.86)
• Cocoon: Robust Multi-Modal Perception with Uncertainty-Aware Sensor Fusion (ICLR 2025 · similarity 0.85)
• UncertainSAM: Fast and Efficient Uncertainty Quantification of the Segment Anything Model (ICML 2025 · similarity 0.85)
• Layer-Wise Modality Decomposition for Interpretable Multimodal Sensor Fusion (NeurIPS 2025 · similarity 0.83)