Deconstructing Stereotypes: Scope-Conditioned Generation for Effective Multilingual Counterspeech
本文提出了一种新的范围条件生成框架，通过将结构化的刻板印象特征整合到大型语言模型提示中，有效生成多语言反言论，以应对在线仇恨言论。
0 citationsRead paper
Selected work
Representative Papers
FoundYou: A Unified Model for Personalized Segmentation and Retrieval
本文提出FoundYou模型，通过利用SAM 2的实例级线索统一解决个性化分割和检索问题，实现两种任务的一体化处理，并在多项基准测试中取得显著性能提升。
0 citationsRead paper
Methods for the Segmentation of Reticular Structures Using 3D LiDAR Data: A Comparative Evaluation
This study addresses the challenge of autonomous navigation for climbing robots on metallic truss infrastructure (e.g., bridges, towers), focusing on semantic segmentation of traversable surfaces in 3D LiDAR point clouds. We propose an analytical algorithm based on eigenvalue decomposition of local planar patches in point clouds and conduct a systematic benchmark against leading deep learning models—PointNet, PointNet++, MinkUNet34C, and PointTransformerV3. Results show that the analytical method achieves near-state-of-the-art accuracy with minimal parameters and high computational efficiency; PointTransformerV3 attains the best performance with 97% mIoU. Crucially, this work is the first to characterize the accuracy–efficiency trade-off between analytical and learning-based approaches on complex, mesh-like truss structures, demonstrating their complementary strengths. The findings establish a deployable pathway toward real-time, robust navigation for robots operating in unstructured metallic truss environments.
0 citationsRead paper
Recent publications
Latest Papers
Deconstructing Stereotypes: Scope-Conditioned Generation for Effective Multilingual Counterspeech
本文提出了一种新的范围条件生成框架，通过将结构化的刻板印象特征整合到大型语言模型提示中，有效生成多语言反言论，以应对在线仇恨言论。
0 citationsRead paper
FoundYou: A Unified Model for Personalized Segmentation and Retrieval
本文提出FoundYou模型，通过利用SAM 2的实例级线索统一解决个性化分割和检索问题，实现两种任务的一体化处理，并在多项基准测试中取得显著性能提升。
0 citationsRead paper
Methods for the Segmentation of Reticular Structures Using 3D LiDAR Data: A Comparative Evaluation
This study addresses the challenge of autonomous navigation for climbing robots on metallic truss infrastructure (e.g., bridges, towers), focusing on semantic segmentation of traversable surfaces in 3D LiDAR point clouds. We propose an analytical algorithm based on eigenvalue decomposition of local planar patches in point clouds and conduct a systematic benchmark against leading deep learning models—PointNet, PointNet++, MinkUNet34C, and PointTransformerV3. Results show that the analytical method achieves near-state-of-the-art accuracy with minimal parameters and high computational efficiency; PointTransformerV3 attains the best performance with 97% mIoU. Crucially, this work is the first to characterize the accuracy–efficiency trade-off between analytical and learning-based approaches on complex, mesh-like truss structures, demonstrating their complementary strengths. The findings establish a deployable pathway toward real-time, robust navigation for robots operating in unstructured metallic truss environments.
0 citationsRead paper