Being a cross-camera retrieval task, person re-identification suffers fromimage style variations caused by different cameras. The art implicitlyaddresses this problem by learning a camera-invariant descriptor subspace. Inthis paper, we explicitly consider this challenge by introducing camera style(CamStyle) adaptation. CamStyle can serve as a data augmentation approach thatsmooths the camera style disparities. Specifically, with CycleGAN, labeledtraining images can be style-transferred to each camera, and, along with theoriginal training samples, form the augmented training set. This method, whileincreasing data diversity against over-fitting, also incurs a considerablelevel of noise. In the effort to alleviate the impact of noise, the labelsmooth regularization (LSR) is adopted. The vanilla version of our method(without LSR) performs reasonably well on few-camera systems in whichover-fitting often occurs. With LSR, we demonstrate consistent improvement inall systems regardless of the extent of over-fitting. We also reportcompetitive accuracy compared with the state of the art.
展开▼
