LiteMapNet: Lightweight Online High-Definition Mapping With Attention Head Pruning

High-definition (HD) maps provide precise geometric priors for perception and planning in autonomous driving; however, existing online HD map construction models are often computationally demanding, which hampers real-time deployment on resource-constrained in-vehicle platforms. To address this issue, we propose LiteMapNet, a lightweight transformer-based framework that accelerates online HD map construction by eliminating redundancy in self-attention and cross-attention modules. LiteMapNet employs a post-training, two-stage structured attention-head pruning pipeline. In the first stage, we estimate attention-head importance using the Fisher information matrix and prune redundant heads with minimal impact on mapping accuracy under a specified compute and latency budget. In the second stage, we introduce continuous learnable mask variables and optimize them through lightweight calibration to recover performance while improving stability and generalization, without incurring additional inference-time overhead. Experiments on nuScenes and Argoverse 2 show that LiteMapNet substantially reduces computational cost and floating point operations using only a small calibration set. While keeping the accuracy drop below 1% as compared to the unpruned MapQR baseline, LiteMapNet achieves up to 1.5× inference speedup on a 4× NVIDIA H100 GPU server, enabling efficient deployment of online HD map models in resource-limited settings.