Pedestrian Alignment Network for Large-scale Person Re-identification

Person re-identification (person re-ID) is mostly viewed as an imageretrieval problem. This task aims to search a query person in a large imagepool. In practice, person re-ID usually adopts automatic detectors to obtaincropped pedestrian images. However, this process suffers from two types ofdetector errors: excessive background and part missing. Both errors deterioratethe quality of pedestrian alignment and may compromise pedestrian matching dueto the position and scale variances. To address the misalignment problem, wepropose that alignment can be learned from an identification procedure. Weintroduce the pedestrian alignment network (PAN) which allows discriminativeembedding learning and pedestrian alignment without extra annotations. Our keyobservation is that when the convolutional neural network (CNN) learns todiscriminate between different identities, the learned feature maps usuallyexhibit strong activations on the human body rather than the background. Theproposed network thus takes advantage of this attention mechanism to adaptivelylocate and align pedestrians within a bounding box. Visual examples show thatpedestrians are better aligned with PAN. Experiments on three large-scale re-IDdatasets confirm that PAN improves the discriminative ability of the featureembeddings and yields competitive accuracy with the state-of-the-art methods.