0 views collected around this technical thread.
The article introduces PreFLMR, an open‑source, general‑purpose pre‑trained multimodal retriever that leverages fine‑grained late‑interaction to boost retrieval‑augmented generation for knowledge‑intensive visual tasks, describes its M2KR benchmark, training stages, and strong experimental results across multiple tasks.
This article details Baidu's multimodal retrieval architecture, explaining the separation of online and offline services, the design of the Imazon image processing and ingestion platform, its technical indicators, large‑scale streaming and batch pipelines, optimization practices for high throughput, and the underlying content‑relationship engine.
This article details Baidu's large‑scale image processing and multimodal retrieval system, describing its offline‑online architecture, massive data ingestion pipeline, ANN search techniques, performance metrics, infrastructure components, and a series of optimizations for throughput, cost, and reliability in a high‑volume streaming environment.
The Imazon platform unifies Baidu’s image acquisition, feature extraction, and ANN‑based multimodal retrieval into a cloud‑native, real‑time pipeline that ingests billions of images daily, optimizes storage and GPU usage, reduces message‑queue costs, and ensures high‑throughput, low‑latency search across text, visual, and voice queries.
The ‘One Name’ team from Nanjing University achieved a MAP of 0.8986 and third place in the iQIYI multimodal video retrieval challenge by fusing official face embeddings with scene features, using channel‑attention‑based video feature fusion, a multimodal SE‑ResNeXt module, and a carefully partitioned model ensemble.