JD ARVR Tech Department Publishes Two Papers on Defocus Blur Detection and Few-Shot Learning in Top Venues

JD.com, an internet company linking production and consumption, emphasizes technology, efficiency, and sustainability as its core mission, leveraging a solid supply‑chain system and extensive data foundation.

Recently, the JD ARVR technology department had two research papers accepted by top venues: one in the SCI‑Q1 journal Transaction on Multimedia (impact factor 5.452) and another at the premier AI conference IJCAI 2020 (12.6% acceptance rate), demonstrating JD’s strength in multimedia and artificial intelligence.

The ARVR team has delivered mature solutions for JD’s core businesses—e‑commerce, advertising, offline retail—and explored innovations in publishing, education, logistics, and smart cities, achieving 100% self‑developed AR try‑on technologies (e.g., virtual makeup, shoe fitting) with an industry‑leading AR engine.

Guided by a PaaS and platform strategy, the department built the “Tian Gong AR Open Platform,” serving as a crucial bridge to partners.

The first paper, BR2Net: Defocus Blur Detection via Bi‑directional Channel Attention Residual Refining Network , addresses the challenge of detecting out‑of‑focus regions caused by camera focal distance mismatches, which is vital for image deblurring, segmentation, depth estimation, object recognition, scene classification, and image quality assessment.

BR2Net employs a deep convolutional neural network with a Residual Learning and Optimization Module (RLRM) to correct prediction errors, bi‑directional residual feature refinement, and channel‑attention mechanisms to enhance discriminative features; experiments show it surpasses current state‑of‑the‑art methods in both speed and accuracy.

The second paper, Transductive Relation‑Propagation Network for Few‑shot Learning , proposes a novel graph‑neural‑network‑based approach that explicitly models and propagates relationships between support‑query sample pairs, introduces a pseudo‑node for query features, and adopts transductive learning to improve few‑shot classification across vision and language tasks.

TRPN’s design yields significant performance gains on multiple benchmark datasets, achieving results comparable to the latest heavyweight models while using a lightweight backbone, thus offering a practical solution for real‑world applications.

Looking ahead, JD’s ARVR department plans to broaden its business coverage, empower more internal and external scenarios with AR, and continue delivering mature solutions and components through a PaaS model, co‑building a 5G‑era ARVR ecosystem with industry partners.