Scholarlink

In the rapidly evolving landscape of multimedia data, efficient content-based video retrieval is increasingly vital. Addressing the HCM AI Challenge 2025, we propose a robust content-based video retrieval pipeline that integrates FastAPI for high-performance APIs, Milvus for scalable vector search, MongoDB for metadata storage, and MinIO for object storage. Our approach employs the laion/CLIP-ViT-B-32-laion2B-s34B-b79K model to generate 512-dimensional keyframe embeddings, stored in PyTorch as a .pth file, enabling precise visual content retrieval. Keyframes are mapped via a converted id2index.json index for efficient ID-to-path resolution. The workflow covers Docker-based environment setup, data conversion, embedding/keyframe migration into Milvus and MongoDB/MinIO, and deployment of a FastAPI backend with a Streamlit interface for interactive querying and visualization. To contextualize retrieval performance, we additionally report a proxy benchmark comparison against representative baselines in the LoVR [3] table. Our scaled proxy score reaches \(R@10 = 96.85\) , substantially exceeding the CLIP baseline (47.11, +49.74) and outperforming stronger image encoders such as MetaCLIP-ViT-H-14 (58.89, +37.96) and EVA02-CLIP-B-16 (48.61, +48.24). These results highlight the effectiveness of our scalable infrastructure design and provide a strong foundation for further optimization toward competitive long-video retrieval systems. Our code implementation is available at: https://github.com/VoThiKimTrang06101997/HCM_AI_Challenge

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Scalable Architecture for Multimodal Video Retrieval Using CLIP Embeddings with the Integration of Milvus Indexing, MinIO Storage, and MongoDB Metadata Fusion

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Thi Kim Trang Vo,
  • Xuan Hieu Tran

摘要

In the rapidly evolving landscape of multimedia data, efficient content-based video retrieval is increasingly vital. Addressing the HCM AI Challenge 2025, we propose a robust content-based video retrieval pipeline that integrates FastAPI for high-performance APIs, Milvus for scalable vector search, MongoDB for metadata storage, and MinIO for object storage. Our approach employs the laion/CLIP-ViT-B-32-laion2B-s34B-b79K model to generate 512-dimensional keyframe embeddings, stored in PyTorch as a .pth file, enabling precise visual content retrieval. Keyframes are mapped via a converted id2index.json index for efficient ID-to-path resolution. The workflow covers Docker-based environment setup, data conversion, embedding/keyframe migration into Milvus and MongoDB/MinIO, and deployment of a FastAPI backend with a Streamlit interface for interactive querying and visualization. To contextualize retrieval performance, we additionally report a proxy benchmark comparison against representative baselines in the LoVR [3] table. Our scaled proxy score reaches \(R@10 = 96.85\) , substantially exceeding the CLIP baseline (47.11, +49.74) and outperforming stronger image encoders such as MetaCLIP-ViT-H-14 (58.89, +37.96) and EVA02-CLIP-B-16 (48.61, +48.24). These results highlight the effectiveness of our scalable infrastructure design and provide a strong foundation for further optimization toward competitive long-video retrieval systems. Our code implementation is available at: https://github.com/VoThiKimTrang06101997/HCM_AI_Challenge