Skip to main content

Cross-Extremely Low-Resolution Re-Identification via Dynamic Multi-Scale Learning

By
Yuxuan Tong; Shiyu Song; Xiaolin Chen; Zhihan Wang; Tianfeng Wang; Jiabao Wang

Existing re-identification methods typically assume that both query and gallery images share similar, high resolutions. However, in real-world scenarios such as UAV reconnaissance and wide-area surveillance, objects often appear as eXtremely Low-Resolution (XLR) instances, for example, pedestrians as small as 16× 8pixels. Such a substantial resolution mismatch causes a dramatic drop in performance. Current super-resolution (SR) based approaches are ill-suited for handling 8× or even 16× magnification factors, and they introduce additional computational cost. To address cross-extremely low-resolution Re-ID (XLR-ReID), we propose a dynamic multi-scale learning approach. It employs the DINOv3 backbone for robust feature extraction, along with a dynamic multi-scale learning module that randomly applies 2×–16× down-sampling to high-resolution images and then up-samples them back to the original size, which is used to force the model to learn degradation-robust features without relying on any SR module. Additionally, a hybrid loss function combining prototypical contrastive loss and identification loss is used to enforce intra-class compactness and inter-class separability. Under 16×8 (pedestrian) and 16×16 (vehicle) queries, experiments show that our method significantly outperforms existing models, improving Rank-1 accuracy by 39.5% for pedestrians (on Market-1501) and 11.4% for vehicles (on VeRi-776) over the best baseline. This provides an effective and efficient solution for XLR-ReID.

Read on IEEE Xplore