Larger Context Window, Fewer Overcorrections: Optimizing Prompts and Batching for Minimal-Edit Grammatical Error Correction
本文通过引入基于分类的指令、批量处理句子和优化提示的方法，解决了大型语言模型在最小编辑语法纠错任务中过度纠正的问题。
0 citationsRead paper
Selected work
Representative Papers
DF26: We Cannot Tell Fake From Real Anymore
本文通过引入DF26基准来解决AI生成视频的检测问题，该基准包含271个真实视频和2420个由现代模型生成的合成视频，揭示了当前检测方法的局限性。
0 citationsRead paper
QCell: Recombining and Aligning Cell Queries for Overlapping Instance Segmentation
QCell通过重组和对齐细胞查询来解决显微图像中重叠细胞实例分割的问题，利用实例重组模块和对比查询对齐目标进行全局推理。
0 citationsRead paper
Synthetic Data Augmentation for Satellite-Based Analysis of Battle-Damaged Agricultural Fields in Ukraine
This study addresses the challenges of scarce annotations and class imbalance in monitoring war-damaged farmland in Ukraine by proposing a synthetic data augmentation framework integrating conditional GANs and DDPMs, coupled with a Vision Transformer for classification. The research validates the efficacy of balanced DDPMs in geospatial few-shot scenarios, significantly mitigating data scarcity bottlenecks. Experimental results demonstrate that the model achieves 88% accuracy and a macro F1-score of 78%, while notably improving the recall for non-bombed areas from 41% to 69%. By establishing an efficient few-shot learning paradigm for war damage assessment, this work offers substantial practical value for real-world applications in conflict-affected agricultural monitoring.
0 citationsRead paper
SLAM in Low-Light Environments: Project Report
This study systematically evaluates six representative SLAM approaches—ORB-SLAM3, DSO, Kimera-VIO, OpenVINS, DPVO, and DPV-SLAM—under low-light conditions using a unified benchmark, the LaMARia dataset. Encompassing feature-based, direct, filtering, and learning-based paradigms, the methods are quantitatively assessed via absolute and relative pose errors as well as control point recall rates to analyze accuracy and robustness. Results reveal that only Kimera-VIO successfully tracks all sequences with the lowest relative error, while DPVO and DPV-SLAM, though frame-loss-free, exhibit absolute trajectory errors on the order of hundreds of meters. Other RGB-only methods generally fail. The findings underscore the critical role of inertial fusion and global optimization in enhancing robustness under poor illumination and delineate the current performance limits of purely visual SLAM in low-light scenarios.
0 citationsRead paper
Recent publications
Latest Papers
Larger Context Window, Fewer Overcorrections: Optimizing Prompts and Batching for Minimal-Edit Grammatical Error Correction
本文通过引入基于分类的指令、批量处理句子和优化提示的方法，解决了大型语言模型在最小编辑语法纠错任务中过度纠正的问题。
0 citationsRead paper
DF26: We Cannot Tell Fake From Real Anymore
本文通过引入DF26基准来解决AI生成视频的检测问题，该基准包含271个真实视频和2420个由现代模型生成的合成视频，揭示了当前检测方法的局限性。
0 citationsRead paper
QCell: Recombining and Aligning Cell Queries for Overlapping Instance Segmentation
QCell通过重组和对齐细胞查询来解决显微图像中重叠细胞实例分割的问题，利用实例重组模块和对比查询对齐目标进行全局推理。
0 citationsRead paper
Synthetic Data Augmentation for Satellite-Based Analysis of Battle-Damaged Agricultural Fields in Ukraine
This study addresses the challenges of scarce annotations and class imbalance in monitoring war-damaged farmland in Ukraine by proposing a synthetic data augmentation framework integrating conditional GANs and DDPMs, coupled with a Vision Transformer for classification. The research validates the efficacy of balanced DDPMs in geospatial few-shot scenarios, significantly mitigating data scarcity bottlenecks. Experimental results demonstrate that the model achieves 88% accuracy and a macro F1-score of 78%, while notably improving the recall for non-bombed areas from 41% to 69%. By establishing an efficient few-shot learning paradigm for war damage assessment, this work offers substantial practical value for real-world applications in conflict-affected agricultural monitoring.
0 citationsRead paper
SLAM in Low-Light Environments: Project Report
This study systematically evaluates six representative SLAM approaches—ORB-SLAM3, DSO, Kimera-VIO, OpenVINS, DPVO, and DPV-SLAM—under low-light conditions using a unified benchmark, the LaMARia dataset. Encompassing feature-based, direct, filtering, and learning-based paradigms, the methods are quantitatively assessed via absolute and relative pose errors as well as control point recall rates to analyze accuracy and robustness. Results reveal that only Kimera-VIO successfully tracks all sequences with the lowest relative error, while DPVO and DPV-SLAM, though frame-loss-free, exhibit absolute trajectory errors on the order of hundreds of meters. Other RGB-only methods generally fail. The findings underscore the critical role of inertial fusion and global optimization in enhancing robustness under poor illumination and delineate the current performance limits of purely visual SLAM in low-light scenarios.
0 citationsRead paper