Adaptive super-resolution for person re-identification with low-resolution images

Ke Han,Yan Huang,Chunfeng Song,Liang Wang,Tieniu Tan
DOI: https://doi.org/10.1016/j.patcog.2020.107682
IF: 8
2021-06-01
Pattern Recognition
Abstract:<p>Person re-identification is challenging with low-resolution query and high-resolution gallery images. To address the resolution mismatch, many methods perform super-resolution (SR) on low-resolution queries with specifying a single scale factor. However, using a single SR module, whichever scale factor is specified, always brings both advantages and drawbacks in recovering and identifying identity information. A larger scale factor recovers more details but produces excessive artifacts, while a smaller one is on the contrary. To exploit their complementary property for more robust recovery and identification, we propose the Adaptive Person Super-Resolution (APSR) model. APSR jointly trains and fuses multiple SR modules based on their generated visual contents, to fully compensate and learn the complementary identity features in an end-to-end manner. To improve the robustness to artifacts during fusion, our model further learns informative features by online dividing and integrating the generated body regions. Extensive experiments verify the effectiveness of our method with state-of-the-art performances.</p>
computer science, artificial intelligence,engineering, electrical & electronic
What problem does this paper attempt to address?