ER-PIQA: A TASK-GUIDED PEDESTRIAN IMAGE QUALITY ASSESSMENT VIA EMBEDDING RECONSTRUCTION

Image quality is an important factor for pedestrian recognition systems. Pedestrian image quality assessment aims at evaluating images in order to provide more reliable and stable images for the following analysis process. Previous work proposed supervised solutions that require artificially or manually labelled quality values. However, due to the lack of a clear quality definition and the variety of recognition tasks, there are still some difficulties in subjective labeling methods. In this paper, a novel task-guided method is proposed to measure pedestrain image quality based on embedding reconstruction without the involvement of subjective labels. Considering the various attention area in different tasks, the pedestrian image quality can be estimated by comparing the similarity between the semantics representation and tasks embedding after reconstruction. Experiments on both person re-identification and pedestrian attribute recognition show advantages of the proposed ER-PIQA. Meanwhile, our approach can be integrated into current recognition systems and adaptively modified for other tasks beyond pedestrian recognition.