人物交互检测（HOI）Human-Object Interaction Detection
论文笔记Paper Notes
HOI 人物交互检测综述：从组合关系到开放词汇HOI Detection Review: From Compositional Relations to Open Vocabulary
按人/物定位、交互分类、长尾泛化、开放词汇和视觉语言模型梳理 HOI 的基本问题与研究路线。A roadmap of HOI detection around localization, interaction classification, long-tail generalization, open vocabulary, and VLMs.
阅读论文笔记Read noteGEN-VLKT: Simplify Association and Enhance Interaction Understanding
用实例级解码简化人—物匹配，并把 CLIP 视觉语言知识迁移到交互理解中。Instance-level decoding with CLIP-based vision-language knowledge transfer for interaction understanding.
阅读论文笔记Read noteCMMP: Conditional Multi-Modal Prompts for Zero-shot HOI Detection
围绕视觉提示和语言提示做条件化设计，提升未见交互类别识别。Conditional visual and language prompts for better zero-shot HOI generalization.
阅读论文笔记Read noteEZ-HOI: Efficient Zero-Shot HOI Detection via VLM Adaptation
把大视觉语言模型的外部知识通过轻量适配注入 HOI 检测。Efficient VLM adaptation for unseen interaction recognition in HOI detection.
阅读论文笔记Read noteSGC-Net: Large Foundation Models for Open-Vocabulary HOI Detection
用大模型生成的细粒度交互描述和分层比较机制支持开放词汇 HOI。Granular LLM-generated interaction descriptions and hierarchical comparison for open-vocabulary HOI.
阅读论文笔记Read note代表工作Selected work
项目经历Project Experience
MAHOI: Multilingual Semantic Recalibration for Open-Vocabulary HOI Detection
以多语言 HOI 描述作为互补语义锚点，设计跨语言上下文交互、动态英语原型库和类别级全局语义约束，提升未见类别的开放词汇 HOI 检测能力。Uses multilingual HOI descriptions as complementary semantic anchors with cross-language interaction, dynamic English prototypes, and category-level constraints for unseen-category generalization.
盖板玻璃缺陷检测Cover Glass Defect Detection
面向高分辨率盖板玻璃质检图像，构建从辅助标注、重叠切块、缺陷分类到分割定位和整图回映射的检测流程。An inspection workflow for high-resolution cover glass images, spanning assisted annotation, overlapped patch sampling, defect classification, segmentation, and full-image remapping.
查看项目View case study仓储图像拼接与模糊复原Warehouse Stitching & Deblurring
面向 AGV 多镜头货架拍摄，开发关键点匹配拼接和一维码去模糊模块，提升成像质量与识别鲁棒性。Built keypoint-based multi-camera stitching and barcode deblurring modules for AGV warehouse imaging and robust recognition.