Selected work

Representative Papers

PseudoMapLabeler: Confidence-Aware Pseudo-Label Generation for Semi-Supervised Online Mapping

Aug 12, 2026

This work addresses the limited generalization capability in high-definition map construction caused by scarce annotated data by proposing a semi-supervised learning approach based on a teacher–student framework. A teacher model trained on a small set of labeled data generates fine-grained pseudo-labels by modeling temporal observation confidence via a Beta distribution and preserving high-confidence regions through a spatial cropping mechanism. These refined pseudo-labels, combined with an optimized map prior, guide the training of the student model. Unlike conventional coarse-grained strategies that discard entire elements, the proposed method retains informative regions, significantly improving performance under low-label regimes. Evaluated on the nuScenes dataset, the approach achieves a 6.1 mAP gain using only minimal labeled data, effectively alleviating dependence on extensive annotations.

0 citationsRead paper

D2HDMap: Non-visible Driveline Map Prior for Online Vectorized HD Map Prediction

Jun 16, 2026

This work addresses the high cost and infrequent update challenges of high-definition (HD) maps, as well as the limited reliability of purely sensor-based online mapping in long-range and occluded scenarios. To this end, the authors propose D2HDMap, a system that leverages a lightweight, imperceptible driveline prior to guide an end-to-end neural network for online vectorized HD map prediction. By incorporating a noise-aware training strategy, the method effectively fuses prior knowledge with real-time perception. Remarkably, using only this low-cost and easily maintainable driveline prior, D2HDMap significantly enhances model generalization even in the absence of such priors during inference. Experimental results demonstrate that D2HDMap achieves 44.8 mAP on geographically disjoint test sets of nuScenes and Argoverse 2, outperforming state-of-the-art methods while exhibiting greater robustness to localization errors.

0 citationsRead paper

Using Language and Road Manuals to Inform Map Reconstruction for Autonomous Driving

Jun 12, 2025

This paper addresses the low accuracy and poor generalization of lane topology prediction in complex intersections for autonomous driving. We propose a lightweight enhancement method that jointly integrates linguistic semantics (e.g., road names) and structured priors (e.g., lane width specifications from roadway design manuals), explicitly encoding them alongside OpenStreetMap (OSM) metadata and SMERF online map priors to enable joint reasoning over linguistic rules and geometric constraints. Our approach incorporates centerline semantic enhancement and employs a topology-aware multi-metric evaluation framework comprising four complementary metrics. Experiments on two geographically diverse complex intersection benchmarks demonstrate significant improvements in lane and traffic element detection, association, and topology prediction accuracy. The results validate the robustness and cross-scene scalability of the proposed method.

0 citationsRead paper

An Inclusive Foundation Model for Generalizable Cytogenetics in Precision Oncology

May 21, 2025

Chromosome abnormality identification is critical for genetic disorder diagnosis and precision oncology, yet existing AI approaches suffer from high annotation costs, incomplete dataset coverage, and poor generalizability. To address these limitations, we propose CHROMA—the first inclusive foundation model designed for comprehensive chromosome abnormality detection across all major types, enabling robust generalization under resource-constrained settings (e.g., few-shot learning and severe class imbalance). CHROMA leverages self-supervised learning, pretrained on 84,000 clinical specimens (~4 million high-resolution karyotype images), and integrates multi-scale feature representation with anomaly-decoupled modeling. Extensive evaluations across diverse chromosome abnormality detection tasks demonstrate consistent and significant superiority over state-of-the-art methods. CHROMA markedly reduces reliance on expert annotations, enhances early detection of rare genomic aberrations, and advances the clinical deployment of scalable, trustworthy AI in cytogenetics.

0 citationsRead paper

SD++: Enhancing Standard Definition Maps by Incorporating Road Knowledge using LLMs

Feb 04, 2025

To address the high cost and restricted accessibility of high-definition (HD) maps, this paper proposes a large language model (LLM)-driven road manual knowledge augmentation method for semantic-degraded (SD) maps. Our approach systematically parses multi-national road manuals (e.g., California, USA; Japan), integrating geospatial prompting with multi-strategy LLM inference (LLaMA/GPT series) to automate extraction and injection of road network geometry and semantic knowledge. This work is the first to synergize formal road regulation documents with LLM reasoning for map enhancement, and introduces a topology-aligned semantic injection framework enabling cross-regional generalization. Experiments on California and Japanese datasets demonstrate that our method reduces average lane centerline error in SD maps by 38% and achieves 92% semantic label completeness—substantially improving lane-level accuracy and practical utility of low-cost SD maps.

0 citationsRead paper

Recent publications

Latest Papers

PseudoMapLabeler: Confidence-Aware Pseudo-Label Generation for Semi-Supervised Online Mapping

Aug 12, 2026

This work addresses the limited generalization capability in high-definition map construction caused by scarce annotated data by proposing a semi-supervised learning approach based on a teacher–student framework. A teacher model trained on a small set of labeled data generates fine-grained pseudo-labels by modeling temporal observation confidence via a Beta distribution and preserving high-confidence regions through a spatial cropping mechanism. These refined pseudo-labels, combined with an optimized map prior, guide the training of the student model. Unlike conventional coarse-grained strategies that discard entire elements, the proposed method retains informative regions, significantly improving performance under low-label regimes. Evaluated on the nuScenes dataset, the approach achieves a 6.1 mAP gain using only minimal labeled data, effectively alleviating dependence on extensive annotations.

0 citationsRead paper

D2HDMap: Non-visible Driveline Map Prior for Online Vectorized HD Map Prediction

Jun 16, 2026

This work addresses the high cost and infrequent update challenges of high-definition (HD) maps, as well as the limited reliability of purely sensor-based online mapping in long-range and occluded scenarios. To this end, the authors propose D2HDMap, a system that leverages a lightweight, imperceptible driveline prior to guide an end-to-end neural network for online vectorized HD map prediction. By incorporating a noise-aware training strategy, the method effectively fuses prior knowledge with real-time perception. Remarkably, using only this low-cost and easily maintainable driveline prior, D2HDMap significantly enhances model generalization even in the absence of such priors during inference. Experimental results demonstrate that D2HDMap achieves 44.8 mAP on geographically disjoint test sets of nuScenes and Argoverse 2, outperforming state-of-the-art methods while exhibiting greater robustness to localization errors.

0 citationsRead paper

Using Language and Road Manuals to Inform Map Reconstruction for Autonomous Driving

Jun 12, 2025

This paper addresses the low accuracy and poor generalization of lane topology prediction in complex intersections for autonomous driving. We propose a lightweight enhancement method that jointly integrates linguistic semantics (e.g., road names) and structured priors (e.g., lane width specifications from roadway design manuals), explicitly encoding them alongside OpenStreetMap (OSM) metadata and SMERF online map priors to enable joint reasoning over linguistic rules and geometric constraints. Our approach incorporates centerline semantic enhancement and employs a topology-aware multi-metric evaluation framework comprising four complementary metrics. Experiments on two geographically diverse complex intersection benchmarks demonstrate significant improvements in lane and traffic element detection, association, and topology prediction accuracy. The results validate the robustness and cross-scene scalability of the proposed method.

0 citationsRead paper

An Inclusive Foundation Model for Generalizable Cytogenetics in Precision Oncology

May 21, 2025

Chromosome abnormality identification is critical for genetic disorder diagnosis and precision oncology, yet existing AI approaches suffer from high annotation costs, incomplete dataset coverage, and poor generalizability. To address these limitations, we propose CHROMA—the first inclusive foundation model designed for comprehensive chromosome abnormality detection across all major types, enabling robust generalization under resource-constrained settings (e.g., few-shot learning and severe class imbalance). CHROMA leverages self-supervised learning, pretrained on 84,000 clinical specimens (~4 million high-resolution karyotype images), and integrates multi-scale feature representation with anomaly-decoupled modeling. Extensive evaluations across diverse chromosome abnormality detection tasks demonstrate consistent and significant superiority over state-of-the-art methods. CHROMA markedly reduces reliance on expert annotations, enhances early detection of rare genomic aberrations, and advances the clinical deployment of scalable, trustworthy AI in cytogenetics.

0 citationsRead paper

SD++: Enhancing Standard Definition Maps by Incorporating Road Knowledge using LLMs

Feb 04, 2025

To address the high cost and restricted accessibility of high-definition (HD) maps, this paper proposes a large language model (LLM)-driven road manual knowledge augmentation method for semantic-degraded (SD) maps. Our approach systematically parses multi-national road manuals (e.g., California, USA; Japan), integrating geospatial prompting with multi-strategy LLM inference (LLaMA/GPT series) to automate extraction and injection of road network geometry and semantic knowledge. This work is the first to synergize formal road regulation documents with LLM reasoning for map enhancement, and introduces a topology-aligned semantic injection framework enabling cross-regional generalization. Experiments on California and Japanese datasets demonstrate that our method reduces average lane centerline error in SD maps by 38% and achieves 92% semantic label completeness—substantially improving lane-level accuracy and practical utility of low-cost SD maps.

0 citationsRead paper