DA3D: Domain-Aware Dynamic Adaptation for All-Weather Multimodal 3D Detection

Project page for DA3D, published at ACM MM 2025.

Haochen Yang1, Lei Li1, Jiacheng Guo1, Baolu Li1, Minghai Qin1,2, Hongkai Yu1, Tianyun Zhang1
1Cleveland State University 2Western Digital Research
Paper DOI Code
DA3D domain-aware dynamic adaptation framework

Abstract

LiDAR-radar fusion improves sensor robustness in adverse weather, but a weather-agnostic detector still suffers from feature-level domain shift. DA3D reframes all-weather detection as a lightweight capacity-allocation problem and adapts a shared detector without duplicating a complete model for each weather condition.

Weather-domain shift and unified versus independent weather models

Method

Domain-Specific LoRACompact weather-conditioned branches modulate features throughout both the 3D and BEV encoders.
Unified InferenceA pretrained weather classifier activates the appropriate branch while retaining one shared backbone.
Dynamic Rank AllocationGradient sensitivity and SVD reconstruction error guide rank pruning and growth under a fixed budget.

Results

At IoU 0.5, DA3D improves AP3D by 4.9 points on RTNH, 3.8 points on 3D-LRF, and 8.1 points on L4DR. The gains hold across radar-only and LiDAR-radar fusion architectures under normal, overcast, fog, rain, sleet, light-snow, and heavy-snow conditions.

DA3D detection results at IoU 0.5 across seven weather conditions

Ablation Study

The key ablation compares DA3D with an independent model, multi-head adaptation, and fixed-rank MTLoRA. DA3D reaches the strongest IoU 0.5 3D performance while using far fewer parameters than the independent model.

Multi-weather modeling comparison on RTNH

BibTeX

@inproceedings{yang2025da3d,
  title     = {DA3D: Domain-Aware Dynamic Adaptation for All-Weather Multimodal 3D Detection},
  author    = {Yang, Haochen and Li, Lei and Guo, Jiacheng and Li, Baolu and Qin, Minghai and Yu, Hongkai and Zhang, Tianyun},
  booktitle = {Proceedings of the 33rd ACM International Conference on Multimedia},
  year      = {2025},
  doi       = {10.1145/3746027.3755708}
}