CVPR Findings2026
Asymmetric Collaborative Distillation for Asymmetric Image Retrieval
Yi Xie, Huaidong Zhang, Xuandi Luo, Yan Zhou, Shengfeng He
Abstract
Asymmetric image retrieval involves heterogeneous networks operating under differing capacities and input resolutions, often dictated by the deployment constraints of real-world platforms. Conventional methods typically constrain the lowresource query network to mimic a high-capacity gallery network, but this one-way distillation overlooks the intrinsic asymmetry of the task: both networks jointly contribute to inference, and their feature spaces must remain compatible despite resolution and capacity gaps. To address this, we propose the Asymmetric Collaborative Distillation (ACD) framework, which reconceptualizes distillation as a bidirectional alignment problem. ACD freezes a well-trained query network and collaboratively optimizes the gallery network to achieve representational compatibility in the collaborative logit space, where semantic structures are more stable and transferable across resolutions. This enables semantically consistent cross-network calibration. To ensure generalization across both symmetric and asymmetric settings, we propose Resolution-Aware Hierarchical Optimization (RAHO) to selectively calibrate the gallery network, assigning adaptive learning rates to hierarchical module groups based on their contribution to cross-resolution performance. Extensive experiments across four standard benchmarks demonstrate that our method consistently improves retrieval accuracy, yielding notable gains such as +2.25% mAP and +3.78% R1 on
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- Decoupled Knowledge Distillation cites Correlation Congruence for Knowledge Distillation
- Logit Standardization in Knowledge Distillation cites Decoupled Knowledge Distillation
- Logit Standardization in Knowledge Distillation cites Correlation Congruence for Knowledge Distillation
Builds on20
- Swin Transformer V2: Scaling Up Capacity and ResolutionZe Liu, Han Hu, Yutong Lin, Zhuliang Yao et al.CVPR 2022 · 2,138 citations
- Decoupled Knowledge DistillationBorui Zhao, Quan Cui, Renjie Song, Yiyu Qiu et al.CVPR 2022 · 835 citations
- Correlation Congruence for Knowledge DistillationBaoyun Peng, Xiao Jin, Dongsheng Li, Shunfeng Zhou et al.ICCV 2019 · 625 citations
- Logit Standardization in Knowledge DistillationShangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang et al.CVPR 2024 · 183 citations
- Learning Student-Friendly Teacher Networks for Knowledge DistillationDae Young Park, Moon-Hyun Cha, Changwook Jeong, Daesin Kim et al.NeurIPS 2021 · 134 citations
Related papers
- Discretization Is Not Always Better: Rethinking Deep Quantization for Asymmetric Image RetrievalXinze Liu, Dayan Wu, Hengjie Zhu, Chenming Wu et al.AAAI 2026
- D3still: Decoupled Differential Distillation for Asymmetric Image RetrievalYi Xie, Yihong Lin, Wenjie Cai, Xuemiao Xu et al.CVPR 2024 · 10 citations
- Contextual Similarity Distillation for Asymmetric Image RetrievalHui Wu, Min Wang, Wengang Zhou, Houqiang Li et al.CVPR 2022 · 34 citations
- A General Rank Preserving Framework for Asymmetric Image RetrievalHui Wu, Min Wang, Wengang Zhou, Houqiang LiICLR 2023
- Learning Compatible Multi-Prize Subnetworks for Asymmetric RetrievalYushuai Sun, Zikun Zhou, Dongmei Jiang, Yaowei Wang et al.CVPR 2025
